Classifying software requirements is a challenging task, as they are typically specified in natural language. In addition, identifying security-related requirements, especially in the early stages, helps integrate these concepts early in the development process and reduces the likelihood of unauthorized access, data breaches, and other security threats. In this paper, we propose a method for classifying security requirements. We fine-tuned several transformer models by exploiting all model weights during the training process and using the Low-Rank Adaptation approach, which provides more efficient and effective ways of adapting pre-trained transformers to specific tasks. In order to optimize their performance, we employed an hyperparameter tuning phase. Furthermore, we compared and analyzed the performance of the two best models on both approaches and provided explainability through three common methods: attention weight visualization, integrated gradients, and SHAP which help us to better understand the decision-making process of our models.

Privacy is one of the most critical issues associated with spreading the Internet of Things and Internet of Everything devices. Over the years, several methods have been introduced to address this phenomenon. In 2017, Google introduced the concept of Federated Machine Learning. This paradigm allows models to be trained collaboratively across multiple decentralized devices or servers, holding local data samples without exchanging them. This approach enhances data privacy and security by ensuring raw data remains on local devices while only model updates are shared and aggregated. This paper presents a privacy-preserving Android malware detector based on Federated Machine Learning. As a first step, we built a dataset comprising over 40,000 Android applications, including trusted and malicious (belonging to 71 malware families) samples. Afterward, we conducted experiments leveraging three different architectures by exploiting the CIFAR-10 and the ImageNet datasets, employing hyperparameters determined through a Grid Search algorithm by exploiting 40 clients. Moreover, the experimental analysis uses two distributions: Independent and identically distributed and non-independent and identically distributed data. To conclude the Federated Machine Learning experiments, we trained models for each architecture, with both weight types and distribution models, by applying the Clipping Norm Aggregator. The results exhibit interesting performances with Independent and identically distributed data, achieving an accuracy of 0.873 without normalization and 0.877 with the Clipping Norm aggregator. However, with non-independent and identically distributed data, the model accuracy equals 0.865 without normalization, 0.864 with the Clipping Norm aggregator using Custom MobileNet 2. In conclusion, to compare Federated Machine Learning with a centralized training approach, we trained several models adopting the same dataset, dataset splitting, and architectures, achieving an accuracy of 0.944 using InceptionV3. The outcomes show that the proposed method can provide engaging performances in privacy-preserving Android malware detection.

Recently, object detection in bioimage analysis has achieved interesting progress through deep learning models like those in the vision transformers, enabling high-accuracy predictions for tasks such as lesion detection in radiology, ophthalmology and organ segmentation. However, applying these models in healthcare raises critical challenges related to privacy and explainability. Centralized training is often infeasible due to strict data-sharing constraints, such as patient privacy regulations, while the black-box nature of modern detectors hinders clinician trust and adoption, particularly in clinical decision-making. To address these issues, we introduce BIOFED, a framework for privacy-preserving and explainable object detection specifically tailored for biomedical imaging. The idea behind BIOFED is to integrate distributed training of object detectors (for instance, the models belonging to the YOLO family) across decentralized clinical or research clients using aggregation strategies like FedAvg and FedProx, ensuring sensitive patient data never leaves the client side. Furthermore, BIOFED incorporates an explainability module based on Class Activation Mapping techniques, by generating heatmaps that highlight bioimage regions, like suspicious lesions in a chest X-ray or retinal fundus image that are most influential for model predictions. This combination of federated object detection and explainability enhances transparency, accountability, and trustworthiness in safety-critical bioimaging scenarios.

The lack of broad, high-quality, labeled datasets and, crucially, stringent data protection laws frequently make it difficult to integrate Deep Learning (DL), with medical imaging. These limitations make it impossible to centralize data, which results in models that don't generalize to other institutions (non-IID data). To address this, we introduce FLIP_VIT, a novel framework leveraging Federated Learning (FL) for explainable and privacy-preserving biomedical imaging. FL is a decentralized system where the training algorithm moves to the data. Hospitals (clients) train a model locally on their private data, and only the model updates (weights or gradients) are securely shared and aggregated by a central server using techniques like Federated Avareging (FedAvg) and Federeted Proximal (FedProx) to create a global model. Patient data never leaves its source, ensuring privacy and compliance.

In the biomedical field, the increased volume of unstructured data is becoming a prominent field of application for natural language processing, particularly for tasks such as sentence classification. Classifying sentences makes it possible to analyze clinical trial reports and perform clinical decision support. This study presents a framework that transforms sentences into word embedding-generated images to perform sentence classification. The proposed approach leverages the strengths of both natural language processing (NLP) and computer vision to perform robust sentence classification and offer interpretability through image processing. Furthermore, the proposed approach utilizes Class Activation Mapping (CAM) to visualize salient features within the generated images.

Large Language Models have revolutionized Natural Language Processing (NLP), displaying optimal capabilities in numerous tasks including summarization, question-answering, and paraphrasing. However, the intricate mechanisms of the models are often not clear, and hence they are increasingly becoming an inhibiting factor in understanding how they act, where they fall short, and where they may pose risks. This paper proposes a framework for fine-tuning LLMs for diverse text-to-text generation tasks and subsequently applying explainability techniques to foster transparency and trust in these complex models.

With the increasingly growing use of Natural Language Processing (NLP) techniques, medical text classification is becoming a prominent field of application. The ability to employ model-based classification to medical abstracts, i.e. brief texts describing the condition of a patient, can potentially offer significant aid to medical experts in the diagnosis process. Transformer-based models have demonstrated significant advancements in NLP tasks. In this paper, we aim to evaluate the performance of four transformer models i.e., BERT, DeBERTa, DistilBERT, and RoBERTa, in classifying medical abstract as one of five possible conditions. By utilizing a labeled dataset consisting of five categories, we conduct a fine-tuning process of each model and finally analyze their performance using several metrics, specifically accuracy, precision, recall, F1-score, and graphical results, particularly confusion matrices and loss graphs. Our findings show comparable performance across models in terms of accuracy, F1-Score, loss, and confusion matrices. Furthermore, we make use of explainability techniques, specifically integrated gradients, to interpret model predictions by analyzing the least and most impactful tokens. The results show criticalities such as dataset imbalance, as well as a common tendency of the models to misclassify.

Access control policies have become a critical issue in the current digital environment, where data breaches and cyber threats are increasingly common. In this context, Large Language Models can simplify the dynamic adaption of access control policies in response to changes due to evolving technology, business processes, or regulatory requirements. This work presents a study that fine-tuned four different LLMs in access control policy classification by comparing two different approaches to train these models. We fine-tuned the presented models normally and used the Parameter Efficient Fine-Tuning approach, a trade-off between performance and efficiency. We trained four different architectures (BERT, GPT-Neo, RoBERTa, and XLNet) and compared their performance on a text classification task to identify ACP from a labeled dataset of 1,522 samples. We achieved the best performance with a maximum accuracy of 0.91 and an F1-score of 0.93, reflecting a strong balance between precision at 0.97 and recall at 0.90. Finally, we employed integrated gradients to explain the predictions of the best-performing model, enhancing the interpretability of the approach.

Currently, deep learning methods such as CNNs and transformers are primarily used for lower limb motion intention(LLMI) recognition on the basis of surface electromyography (sEMG) signals. However, these methods do not fully and effectively extract the unique spatiotemporal features possessed by sEMG signals. To address this issue, this paper proposes a new model called M2IRV-Net (interpretable Res-ViTNet with multiview interaction and multiobjective optimization). This model introduces a multiview shallow extraction module and combines a parallel architecture of ResNet and Vision Transformer(ViT), leveraging multiview interaction to fully mine key information from different perspectives. Coupled with a multiobjective optimization strategy, the model balances the learning capabilities of various subnetworks, enhancing the accuracy and reliability of lower limb motion intention recognition. Specifically, M2IRV-Net first extracts shallow features through the multiview module, capturing key information from different perspectives. Subsequently, ResNet and ViT process these features in parallel, optimizing them from the perspectives of deep feature representation and global dependency modeling, respectively. Ultimately, the overall model performance is further improved through a multiobjective optimization strategy, ensuring the accuracy and generalizability of motion-intention recognition. To verify the effectiveness of this algorithm, experiments were conducted on the SIAT_LLMD dataset, and the model's performance was further evaluated through ablation experiments. The results indicate that the lower limb motion intention classification accuracy of the M2IRV-Net model is 97.92%, demonstrating the good classification performance of this method.

Breast density classification is a critical factor in assessing breast cancer risk, and most existing methods rely on mammography. This study proposes a method for automatic breast density classification using T2-weighted MRI images, following the ACR BI-RADS criteria. The proposed method involves creating total and border masks based on MRI scans, excluding the chest area, computing the percentage of periglandular fat, and classifying density into “Dense” and “Not Dense”. Experimental validation was conducted using 136 MRI exams, including cases with prosthetic implants, mastectomy, nodulectomy, and tumorectomy. The method achieved an Accuracy of 0.86, a Precision of 0.96, a Specificity of 0.97, and a Sensitivity of 0.76. These results highlight the robustness of the method and demonstrate the potential of MRI-based density classification as a complementary tool to mammography. The findings provide new insights for breast imaging practices and assist radiologists in clinical practice. Future developments will focus on extending the classification to the original four BI-RADS classes. • Breast density is a key risk factor for breast cancer, typically assessed with mammography. • This paper proposes an MRI-based approach for breast density classification into “Dense” and “Not Dense” categories. • The proposed method, based on the computation of periglandular fat percentage, achieved Accuracy 0.87, Precision 0.96, Specificity 0.97, and Sensitivity 0.76

Breast cancer is a kind of cancer that forms in the cells of the breasts. After skin cancer, breast cancer represents the most common cancer diagnosed in women in the United States. As a matter of fact, in January 2022, there are more than 3.8 million women with a history of breast cancer in the United States, this is the reason why there is a need for novel methods for automatic breast cancer screening, with the aim of starting any therapy as quickly as possible to try to limit the proliferation of the disease. In this article, we propose a method aimed at detecting breast cancer through a deep learning network developed by authors. Moreover, the proposed method is able to provide prediction explainability by means of class activation mapping, aimed to automatically highlight the suspicious area on the image. We take into account a way to understand whether the cancer prediction and localization can be considered robust by analyzing the output of two different class activation mapping algorithms. We evaluate the effectiveness of the proposed method by using a dataset composed of 9,016 images obtaining an accuracy equal to 93.5%, thus showing the effectiveness of the proposed network for breast cancer detection and localization.

Skin lesions are any abnormal growths or appearances on the skin, ranging from benign (i.e., non-cancerous) to malignant (i.e., cancerous). The identification of a skin lesion is a crucial task that is carried out in short periods of time to initiate an eventual therapeutic treatment. In this paper, we propose a method for automatic skin lesion detection, implementing Convolutional Neural Networks. Moreover, with the aim of providing a rationale behind the model prediction, we also consider explainability by adopting two different Class Activation Mapping algorithms, which highlight regions in skin images that most contribute to the network’s classification decision. We also include the indices of similarity for further quantitative analysis. Several Convolutional Neural Networks are considered, by obtaining the best results with the MobileNet model, achieving an accuracy equal to 0.935 in skin lesion detection. Moreover, in the experimental analysis, we discuss the effectiveness of Class Activation Mapping algorithms exploited for skin lesion localization.

Malware analysis researchers are currently focused on the design and development of innovative approaches to detect zero-day malware, with particular regard to mobile environments. It is widely recognized that existing free and commercial anti-malware tools struggle to detect unknown threats, operating within the constraints of the signature-based paradigm. Among the set of techniques frequently exploited by researchers for spotting malware, machine learning, with particular regard to deep learning, is quickly gaining prominence as a highly promising method for zero-day malware detection. Unfortunately, many of the proposed methods are not effectively implemented in research prototypes and, therefore, do not truly become usable by end users. For this reason, in this paper, we present the DroidDefender proof-of-concept, an Android antimalware specifically designed and developed to detect zero-day malware. DroidDefender is based on representing an Android application as an image and employs a deep learning model. The DroidDefender research prototype is freely available for research purposes at the following URL: https://www.cybersecurityosservatorio.it/Services/malwareImage.jsp?lang=en.

The ability to manage and enforce security policies in a rapidly changing digital environment is vital. When enforcing and verifying policies on complex systems, it can be challenging for them to be understood by humans. Furthermore, updating access control rules or configurations that are translated from natural language documents might take considerable time and eventually cause errors and damage. It is crucial for the rules and configurations to be modified with care in order to prevent such mistakes. The aim of the following proposal is the design of a method to convert security policies written in natural language into a machine-understandable format that can easily be enforced, verified as well and understood by machines. To this end, we propose using advanced policy languages like XACML and Rego to translate security policies into formats that are automatically enforceable and verifiable by the system. In addition, this transformation will facilitate the implementation of security policies within modern application architectures including micro-services and cloud-native environments. Therefore, under its auspices, a framework that eases policy enforcement and minimizes human errors while ensuring conformity to safety standards set out shall be born out of this research proposal.

In recent years, we have seen an impressive expansion of a family of Artificial intelligence models, known as generative AI, that are capable of producing fresh, unique material, including text, images, audio, and even code. Large datasets of previously published information are used to train these models, enabling them to mimic the patterns and structures of the data and produce original output that is stylistically and qualitatively comparable to the training set. While these models have many promising applications, they also carry significant risks and potential dangers that must be carefully considered, such as misinformation, intellectual property violations, or biased information. For these reasons, in this work, we have proposed a method to detect whether a sentence is human-generated or AI-generated. To achieve this goal, we used a labeled dataset to train four different models from the BERT family, achieving an Accuracy of 96%.

The rapid evolution of deep learning techniques, applied in the smart cities context, has revolutionized computer vision applications, with particular significance in the field of object detection. In this paper, we explore the application of deep learning for real-time car detection in visual data, such as images or video streams. The proposed deep learning model is trained on large-scale datasets containing diverse images of cars, encompassing various lighting conditions, weather patterns, and traffic scenarios. Moreover, once the presence of car(s) is detected, we consider finding the distance between the detected car(s) and the camera: in this way, it is possible to understand the distance to the vehicle to take a certain countermeasure (for example braking or slowing down). The experimental analysis, performed on a dataset composed of 25532 different images confirms the effectiveness of the proposed method for real-world car detection. Moreover, examples of the distance computed between the camera and the detected cars are shown to provide samples of the adoption of the proposed method in the real-world environment.

The detection of road signs is pivotal for effective traffic management and the operation of autonomous vehicles. In this context, deep learning, especially with respect to object detection models, can make a substantial contribution by facilitating the direct detection of road signs from images captured by moving vehicles. In this paper, we propose a method aimed at detecting road signs by exploiting the You Only Look Once model. We experiment with a dataset composed by 10000 different road sign images by obtaining interesting performances, thus showing that the proposed model can be considered reliable for real-time road sign detection.

Ocular diseases present a significant public health challenge, necessitating accurate and timely diagnosis for effective management and preservation of vision. Fundus imaging, providing a single snapshot of the eye’s posterior segment, plays a pivotal role in diagnosing ocular diseases. However, the interpretation of fundus images is inherently limited by the lack of volumetric data, posing challenges for comprehensive disease evaluation. This paper explores a preliminary approach integrating radiomics and pattern matching into the analysis of eye fundus images to enhance diagnostic capabilities. Our method takes as input the radiomic feature values extracted from the medical images and after the selection and discretization process identifies if there exists a recurrent pattern to distinguish diseased eyes from normal ones. We prove our approach on a dataset of 300 eye fundus images. Results achieved are comparable with ones obtained using artificially intelligence-based approaches.

Artificial intelligence is pervading our lives. As a matter of fact, from voice assistants to self-driving cars, a multitude of aspects of our lives are influenced from decisions made by systems trained through artificial intelligence techniques, for instance machine learning. A multitude of decision support systems has been developed as black boxes, i.e. systems that hide the internal logic to the final user. This lack of explanation represents a practical but also ethical issue. Considering the critical issues of the decisions made by these systems in recent times is emerging the need for artificial intelligence techniques able to explain the reasons why a certain decision is taken by these artificial intelligence algorithms. We propose an approach considering the adoption of the counterexample, automatically provided by model checking techniques, with the aim to explain the rationale behind a decision made by the system. Moreover, we propose a case study showing how the proposed approach can be useful to provide explainability in classification tasks.

Temporal Lobe Epilepsy (TLE) is the most common type of focal epilepsy, characterized by abnormalities in brain metabolism usually detected by Magnetic Resonance Imaging (MRI). However, very often additional imaging investigations are required, i.e. Positron Emission Tomography (PET), which is not always available and is invasive for patients. For these reasons, the purpose of this work is to predict the outcome of PET imaging, namely the presence or absence lof focal epilepsy, by using only MRI-negative series with the application of Radiomics and Formal Methods. We collected a series of epilepsy patients who were negative on MRI but positive on PET scans that revealed the presence of abnormalities in brain metabolism, confirming the diagnosis of TLE. From the high-resolution MRI scan, we extracted radiomic features of first and second order. The dataset is exploited through the mathematical logic of Formal Methods, which can manage even small datasets with less impact on robustness. In addition, we studied the effectiveness of Gaussian filtering on MRI images in a pre-processing step, evaluating whether the performance increased. Among all the radiomic features extracted, the Gray Level Cooccurrence Matrix class achieved the best results with $\mathbf{8 6 \%}$ Accuracy, $\mathbf{8 2 \%}$ Precision, $\mathbf{1 0 0 \%}$ Recall and $\mathbf{6 0 \%}$ Specificity, coherently with the state of the art. Concerning the filtering in the Pre-processing step, instead, it can improve diagnosis accuracy in 3 radiomic classes. This methodology would make it possible to assist the medical specialist with a “second-virtual opinion” without any further imaging, improving both the time and costs of the diagnosis.

Background and Objective: Alzheimer’s disease is nowadays the most common cause of dementia. It is a degenerative neurological pathology affecting the brain, progressively leading the patient to a state of total dependence, thus creating a very complex and difficult situation for the family that has to assist him/her. Early diagnosis is a primary objective and constitutes the hope of being able to intervene in the development phase of the disease. Methods: In this paper, a method to automatically detect the presence of Alzheimer’s disease, by exploiting deep learning, is proposed. Five different convolutional neural networks are considered: ALEX_NET, VGG16, FAB_CONVNET, STANDARD_CNN and FCNN. The first two networks are state-of-the-art models, while the last three are designed by authors. We classify brain images into one of the following classes: non-demented, very mild demented and mild demented. Moreover, we highlight on the image the areas symptomatic of Alzheimer presence, thus providing a visual explanation behind the model diagnosis. Results: The experimental analysis, conducted on more than 6000 magnetic resonance images, demonstrated the effectiveness of the proposed neural networks in the comparison with the state-of-the-art models in Alzheimer’s disease diagnosis and localization. The best results in terms of metrics are the best with STANDARD_CNN and FCNN with accuracy, precision and recall between 98% and 95%. Excellent results also from a qualitative point of view are obtained with the Grad-CAM for localization and visual explainability. Conclusions: The analysis of the heatmaps produced by the Grad-CAM algorithm shows that in almost all cases the heatmaps highlight regions such as ventricles and cerebral cortex. Future work will focus on the realization of a network capable of analyzing the three anatomical views simultaneously.

Grading laryngeal squamous cell carcinoma (LSCC) based on histopathological images is a clinically significant yet challenging task. However, more low-effect background semantic information appeared in the feature maps, feature channels, and class activation maps, which caused a serious impact on the accuracy and interpretability of LSCC grading. While the traditional transformer block makes extensive use of parameter attention, the model overlearns the low-effect background semantic information, resulting in ineffectively reducing the proportion of background semantics. Therefore, we propose an end-to-end network with transformers constrained by learned-parameter-free attention (LA-ViT), which improve the ability to learn high-effect target semantic information and reduce the proportion of background semantics. Firstly, according to generalized linear model and probabilistic, we demonstrate that learned-parameter-free attention (LA) has a stronger ability to learn highly effective target semantic information than parameter attention. Secondly, the first-type LA transformer block of LA-ViT utilizes the feature map position subspace to realize the query. Then, it uses the feature channel subspace to realize the key, and adopts the average convergence to obtain a value. And those construct the LA mechanism. Thus, it reduces the proportion of background semantics in the feature maps and feature channels. Thirdly, the second-type LA transformer block of LA-ViT uses the model probability matrix information and decision level weight information to realize key and query, respectively. And those realize the LA mechanism. So, it reduces the proportion of background semantics in class activation maps. Finally, we build a new complex semantic LSCC pathology image dataset to address the problem, which is less research on LSCC grading models because of lacking clinically meaningful datasets. After extensive experiments, the whole metrics of LA-ViT outperform those of other state-of-the-art methods, and the visualization maps match better with the regions of interest in the pathologists' decision-making. Moreover, the experimental results conducted on a public LSCC pathology image dataset show that LA-ViT has superior generalization performance to that of other state-of-the-art methods.

With the growth of the mobile market, malicious applications represent a risk to the security of the users. To mitigate this aspect, researchers proposed different techniques to spot and identify unsafe software placed on the market. On the other hand, malicious writers started to develop ever more sophisticated strategies to hide malicious payloads, in particular through the adoption of obfuscation techniques. The latter consists of hiding the behavior and purpose of malware from antimalware. In this paper, we propose and design a method aimed to detect obfuscated malware. The proposed method builds images directly from system call traces obtained from legitimate, malicious, and obfuscated Android applications. In addition, to show that dynamic analysis and deep learning can build resilient models we propose two experiments using a convolutional neural network. In the first experiment, we train and test the model using a dataset composed of malware, while in the second we train the model using the malware dataset but the model is evaluated using a dataset composed of obfuscated malware. Finally, we analyze the malware and obfuscated detection models from the point of view of explainability using two different class activation mapping algorithms, to understand whether the model predictions can be considered resilient.

In the last two decades, one of the most important topics of discussion is the environmental problem, and to solve the latter, scientists have proposed several types of solutions. Recently Smart Grid concept has taken hold. The latter was presented in 2007 to manage electricity and integrate renewable energy sources. In that context, one of the problems is obtaining power system stability. This pertains to the ability of an electrical grid, transforming smart grid technologies, to maintain steadfast and dependable operation despite the inherent variability introduced by renewable resources. In this paper, we propose a technique to determine if a network is stable or unstable by employing several deep-learning architectures. In detail, we used a dataset retrieved from Kaggle, where all data have been converted into images thanks to the usage of a Python script designed and built by the authors. After getting the related pictures, the new dataset obtained was divided into 80-10-10 (training-validation-test). After this step, we executed several experiments where we reached an accuracy and precision of 0.998. In particular, to perform them we employed different Convolutional Neural Networks present in the state-of-the-art like AlexNet, DenseNet, LeNet, ResNet50, Standard CNN, VGG16, and VGG19. In the final step, to perform the explainability, we also employed three different CAMs Grad-CAM, Grad-CAM++, and Score-CAM, to highlight the most distinctive areas of an image.

Artificial intelligence has increasingly pervaded both everyday life and business environments, assuming a role in supporting human decision-making. These systems have grown progressively intricate and effective, carrying the potential to unearth valuable insights across various applications. Nevertheless, for Artificial Intelligence to gain widespread acceptance, human trust in its outcomes is essential. Trustworthy Artificial Intelligence is a term used to describe legal, ethically compliant, and technically sound Artificial Intelligence. In this paper, we design a proposal with the aim to introduce trustworthiness in Artificial Intelligence models. In the proposed workflow, we consider from a set of models contained in a model repository, the computation of a trustworthy index obtained from the prediction explainability in both training and testing phases. The latter is to consider the most reliable model from the trustworthy index point of view.

Data poisoning represents a set of techniques aimed at perturbing data for training machine learning models, affecting performance. Such intentional attacks are widespread in many applications involving deep learning algorithms and are aimed to provide misclassifications. In this paper, data poisoning on retinal images for the diabetic retinopathy binary classification (health and sick) is presented and evaluated. The presented attacks are almost imperceptible perturbations of the images that nevertheless decrement the metrics of the trained models. Once exposed to the data poisoning distortions on these images, a possible countermeasure to enhance the security from these attacks is shown. In this way, the robustness and vulnerabilities of the network are highlighted and the best result is also analyzed through the use of heatmaps, for the qualitative point of view. The paper aims to focus on the effects of data poisoning in deep learning model testing phase and to discuss possible countermeasures.

Background and objective: The images produced by radiological exams are in DICOM format, the medical imaging standard that includes a large amount of information about the exam and the patient. In a Computed Tomography exam there can be more than 300 DICOM images per patient: this means that Radiomics could addresses the problem of being time and resource-consuming. The research question is: reducing images can lead to improved diagnostic performance?Methods: The study was focused on the classification of healthy, COVID-19 and Lung disease patients. For each patient, a few central images from the Computed Tomography exam are selected, where it is assumed that there is the greatest likelihood of seeing the disease marks. Consequently, this study provides the use of formal and mathematical techniques in order to obtain an early and robust automated diagnosis of COVID-19. In the first phase, the classification is focused on healthy patients, while for COVID-19 patients there is a second phase. The verification will be done on the entire CT examination and then on the CT examination with central selected slices.Results: Results on the original exam are no different from those on the reduced exam on the first phase, and this could an advantage due to the robustness of Formal Methods. In the second phase of the study, almost all indexes become underperforming.Conclusions: A reduced quantity of medical images can be an advantage for medical doctors, but it is not beneficial for diagnosis, as performance remains unchanged or worsens. Indeed, it can be said that disease-relevant information is also taken in the first or last parts of the radiological exam, leading to the claim that Radiomics "sees what the human eye does not see".

In the smart grid ecosystem, the automated reading of water meters is a critical task that can be accomplished with deep learning algorithms instead of the traditional way that relies on the physical presence of employees or utility representatives. The YOLO (You Only Look Once) family of models has gained prominence among these approaches because of its excellent accuracy and quick processing time. In this study, we compare the performance of four YOLO models for automated reading in water metering systems (YOLOv5, YOLOv6, YOLOv7, and YOLOv8). We assess the model's precision, recall, and processing time using a custom dataset of water meter images. In terms of precision and processing time, our testing results demonstrate that YOLOv8 surpasses YOLOv7, YOLOv6, and YOLOv5 by scoring 0.889 in precision, making it a suitable alternative for automated reading in smart metering systems. This benchmarking study provides a valuable reference for researchers and practitioners in the field of smart grid and deep learning, and it can aid in the selection of the most appropriate model for this task.

In recent times is increasing the number of cyberattacks. Daily, we employ computers and mobile devices to perform different tasks, and malicious users are always ready to perpetrate malicious actions. Usually, they use malicious applications able to gain personal information. To avoid these troubles researchers are always focused on the design of new ways to detect malware, typically exploiting machine learning. The progress and recent increasing adoption of quantum computing allowed the inclusion of quantum algorithms into machine learning methods. In this paper, we present a malware detection method based on a quantum machine learning model. In addition, to understand the explainability we adopted two different Class Activation Mapping algorithms to highlight the areas of the images under study that are responsible for a specific categorization of the malware's families. Moreover, we also compare the best results obtained using the quantum model we propose with one of the first convolutional neural network models, i.e., the Le-Net one, to understand whether quantum models can be promising in the malware detection task.

The COVID-19 pandemic has radically changed our daily habits. One of these habits introduced is the use of a face mask, to avoid the spread of the infection, to be used especially in closed and crowded environments. To protect people and to counter the spread of COVID-19, in this paper we propose a method to automatically verify from images whether people are wearing a face mask. We designed a deep learning network to classify whether an image under analysis is with a mask or without a face mask. With the aim to provide explainability to the classifier decision, the proposed method is able to localise the areas of the image under analysis symptomatic of a certain prediction: in this way it is possible to understand the reason why the model predicts a certain label. The experimental analysis considers 7553 (with mask and without mask) images showing that the proposed approach is able to obtain interesting performances in face mask detection.

Detecting structural damage is a fundamental process for visually assessing the quality of any building, such as a bridge or skyscraper. Unfortunately, access to inspect buildings, especially in the case of very tall buildings, is usually not immediate, and in any case, it is difficult. For this reason, unmanned aerial vehicles and drones are used to access inaccessible places, having the possibility of reaching them, making videos and taking photos which are then analyzed by ground personnel. In this paper, we propose a method for the automatic detection of structural damage from images. The proposed method employs an object detection model, i.e., YOLO, for the automatic localization of structural damages on different types of buildings. In the evaluation of a dataset composed of more than 700 images acquired from unmanned aerial vehicles and drones, the proposed method obtains a precision of 0.897, a recall equal to 0.904, and a mean Average Precision value (with an Intersection over Union greater than 0.5) equal to 0.924, showing the effectiveness of the proposed method for the identification and location of structural damage.

Object detection is one of the crucial tasks that has made deep learning of fundamental importance in last years, also thanks to the use of drones and unmanned aerial vehicles able to obtain images and videos in real-time from any location. In the absence of daylight or artificial light for object detection, it is necessary to resort to thermal images, obtained by converting infrared radiation (i.e., heat) into visible images that depict the spatial distribution of temperature differences in a scene viewed by a thermal camera. However, object detection in this type of image and video stream is still challenging due to the complicated scene information and coarse resolution compared to a visible image or video. In this paper, we propose a method aimed to detect objects in thermal images, in particular, the proposed method is aimed to identify persons and dogs acquired from a thermal camera installed, for instance, on drones or unmanned aerial vehicles. We employ an object detection model, i.e., the ”You only look once” one, for the automatic localization of objects in thermal images. In the evaluation of a dataset composed of 203 images with 257 annotations, the proposed method obtains a precision of 0.897, a recall equal to 0.904, and a mean Average Precision value (with an Intersection over Union greater than 0.5) equal to 0.924, showing the effectiveness of the proposed method for the identification and location of persons and dogs from images directly acquired with thermal cameras.

In light of the growing reliance on digital technology, the security of digital devices and networks has become a critical concern in the information technology industry. Network analysis can be helpful for identifying and mitigating network-based attacks, as it enables the monitoring of network behavior and the detection of anomalous activity. Through the use of network analysis, organizations can better defend against potential security threats and protect their interconnected digital systems.

In last years, we are witnessing a growing interest in the application of supervised machine learning techniques in the most disparate fields. One winning factor of machine learning is represented by its ability to easily create models, as it does not require prior knowledge about the application domain. Complementary to machine learning are formal methods, that intrinsically offer safeness check and mechanism for reasoning on failures. Considering the weaknesses of machine learning, a new challenge could be represented by the use of formal methods. However, formal methods require the expertise of the domain, knowledge about modeling language with its semantic and mathematical rigour to specify properties. In this article, we propose a novel learning technique based on the adoption of formal methods for classification thanks to the automatic generation both of the formula and of the model. In this way the proposed method does not require any human intervention and thus it can be applied also to complex/large datasets. This leads to less effort both in using formal methods and in a better explainability and reasoning about the obtained results. Through a set of case studies from different real-world domains (i.e., driver detection, scada attack identification, arrhythmia characterization, mobile malware detection, and radiomics for lung cancer analysis), we demonstrate the usefulness of the proposed method, by showing that we are able to overcome the performances obtained from widespread classification algorithms.

The tumor grading of laryngeal cancer pathological images needs to be accurate and interpretable. The deep learning model based on the attention mechanism-integrated convolution (AMC) block has good inductive bias capability but poor interpretability, whereas the deep learning model based on the vision transformer (ViT) block has good interpretability but weak inductive bias ability. Therefore, we propose an end-to-end ViT-AMC network (ViT-AMCNet) with adaptive model fusion and multiobjective optimization that integrates and fuses the ViT and AMC blocks. However, existing model fusion methods often have negative fusion: 1). There is no guarantee that the ViT and AMC blocks will simultaneously have good feature representation capability. 2). The difference in feature representations learning between the ViT and AMC blocks is not obvious, so there is much redundant information in the two feature representations. Accordingly, we first prove the feasibility of fusing the ViT and AMC blocks based on Hoeffding's inequality. Then, we propose a multiobjective optimization method to solve the problem that ViT and AMC blocks cannot simultaneously have good feature representation. Finally, an adaptive model fusion method integrating the metrics block and the fusion block is proposed to increase the differences between feature representations and improve the deredundancy capability. Our methods improve the fusion ability of ViT-AMCNet, and experimental results demonstrate that ViT-AMCNet significantly outperforms state-of-the-art methods. Importantly, the visualized interpretive maps are closer to the region of interest of concern by pathologists, and the generalization ability is also excellent. Our code is publicly available at https://github.com/Baron-Huang/ViT-AMCNet.

Mobile malware are increasing their complexity to be able to evade the current detection mechanism by gathering our sensitive and private information. For this reason, an active research field is represented by malware detection, with a great effort in the development of deep learning models starting from a set of malicious and legitimate applications. The recent introduction of quantum computing made possible quantum machine learning i.e., the integration of quantum algorithms within machine learning algorithms. In this paper, we propose a comparison between several deep learning models, by taking into account also a hybrid quantum malware detector. We explore the effectiveness of different architectures for malicious family detection in the Android environment: LeNet, AlexNet, a Convolutional Neural Network model designed by authors, VGG16 and a Hybrid Quantum Convolutional Neural Network i.e., a model where the first layer is a quantum convolution that uses transformations in circuits to simulate the behavior of a quantum computer. Experiments performed on a real-world dataset composed of 8446 Android malicious and legitimate applications allow us to compare the various models, with particular regard to the quantum model concerning the other ones.

Cardiovascular diseases include a very long series of diseases that afflict many people in the world. Many of them can be diagnosed by listening to the heartbeat, however in the face of the large number of patients performing checks, great delays can occur given the few doctors available. In this paper we propose a convolutional neural network aimed to discriminate regular heartbeats from abnormal ones, making a first screening of patients. Moreover we provide classification explainability through activation maps. The experimental analysis consists of 3240 (regular and abnormal) heartbeat phonocardiogram signals, showing the effectiveness of the proposed method.

Cervical cancer develops in the lower part of the uterus, the organ of the female apparatus where the embryo is received and develops during pregnancy. In this paper we investigate the possibility to automatically detect the presence of cancerous cells and to predict of the stage of the cancerous lesion of the uterine cervix by exploiting images of cervical cells captured by the microscope. We extract a set of numerical features from each images and we build supervised machine learning models to diagnose the cervix cancer. The experimental analysis show that the proposed method is promising in distinguish between healthy and cancerous cells and to detect also high and low-grade squamous intraepithelial lesions.

Over the years, the number of compromised accounts dramatically increased. To avoid this type of attack various types of authentication methods were introduced. In particular, currently, researchers are focusing on biometric-based techniques such as physical-biometric and behavior-biometric. The idea at the bottom of the last technique is that each person exhibits a unique behavior. Starting from the touch dynamics, and keyboard dynamics nowadays, one of the most promising investigation areas is currently represented by mouse dynamics. Because of the simpler technology necessary to gather biometric data without employing user sensor data, the latter has recently been a popular study area. In this paper, we propose an approach for continuous and silent user authentication based on mouse dynamics and explainable deep learning. We build a set of images using an existing dataset of mouse dynamics in CSV format. The images obtained were then used to train a deep-learning model to discriminate between legitimate and malicious users. We also adopted the Gradient-weighted Class Activation Mapping, to allow highlighting the areas of the images which are responsible for a specific legitimate/attack prediction, thus providing explainability behind the model classification. The preliminary experimental analysis based on ten different users shows that the proposed method can be promising in silent and continuous user authentication.

Many tasks that require a big workforce are automated. In many areas of the world, the consumption of utilities such as electricity, gas, and water is monitored by meters that need to be read by humans. The reading of such meters requires the presence of an employee or a representative of the utility provider. Automatic meter reading is a crucial in the implementation of smart grids. For this reason, with the aim to boost the implementation of the smart grid paradigm, in this paper we propose a method aimed to automatically read digits in a dial meter. In detail the proposed method is aimed to localise the dial meter from an image, to detect the digits and the classify the digits. Deep learning is exploited, in particular the YOLOv5s model is considered for the localisation of digits and for their recognition. An experimental real-world case study is proposed to confirm the effectiveness of the proposed method for automatic digits localisation recognition in dial meter.

Nowadays pathologists have to analyse blood cells manually with the aim to diagnose diseases. In order to perform this manual task blood samples must be collected from the patient and then placed on a microscope slide. This slide is studied with the aim to detect the abnormality presence. The manual method used to identify abnormality of blood cells is tedious, prone to human errors and time consuming. Hence there is a need for computer aided system which can analyze blood cells automatically at faster rate with accuracy. Such systems can be designed using image processing techniques. To automatise this process by helping pathologists, in this paper we propose the adoption of deep learning to automatically count and localise red blood cells, white blood cells and platelets by analysing blood microscopic images. We resort to the YOLO object detection model, able to look at the whole image so its predictions are informed by global context in the image. To show the effectiveness of the proposed method we evaluate our model on a dataset composed by 874 microscopic blood images, obtaining interesting results. Furthermore we show several examples related to how the proposed method can be helpful for the pathologists in their real-world work.

Prostate cancer is classified into different stages, the so-called Gleason scores. The Gleason score assigning represents a time-consuming task usually performed by radiologists. In this paper we propose an approach focused on the automatic Gleason score detection, by exploiting a set of 71 radiomic features directly obtained from magnetic resonance images. The radiomic feature set is directly obtainable from segmented magnetic resonance images. We build several models considering fuzzy supervised machine learning techniques, obtaining with the Fuzzy Unordered Rule Induction Algorithm classification algorithm a precision ranging from 0.865 to 0.953 and a recall from to 0.827 to 0.973. Moreover we show also how the rules automatically obtained from the algorithm can be exploited in order to provide explainability behind the model prediction for the automatic Gleason score diagnosis.

Often when we have a lot of data available we can not give them an interpretability and an explainability such as to be able to extract answers, and even more so diagnosis in the medical field. The aim of this contribution is to introduce a way to provide explainability to data and features that could escape even medical doctors, and that with the use of Machine Learning models can be categorized and "explained".

Deep Learning models demonstrated high accuracies performance in malware classification, but they are still lacking "explainability" to ensure robustness and reliability in the generated prediction. In this short contribution, we summarize the researches that we conducted in the latest years in the Malware Analysis field.

The new coronavirus disease (COVID-19), declared a pandemic on 11 March 2020 by the World Health Organization, has caused over 6 million victims worldwide. Because of the rapid spread of the virus, with the aim to perform screening we exploit deep learning model to quickly diagnose altered respiratory conditions. In this paper, we propose a method to recognize and classify cough audio files into three classes to distinguish patients with COVID-19 disease, symptomatic ones and healthy subjects, with the use of a convolutional neural network (CNN). Cough audios were recorded by using a smartphone and its built-in microphone. From cough recordings, we generate spectrogram images and we obtain an accuracy equal to 0.82 with a deep learning network developed by authors. Our method also provides heatmaps, which show the relevant input areas used by the model for the final forecast, and this aspect ensures the explainability of the method.

Over the years the number of fraud credential attacks increased drastically. To avoid this type of attack many types of authentication methods have been introduced, in particular, recently taken hold biometric-based such as physical-biometric and behavior-biometric. The idea at the bottom of the latter is that each person has a unique behavior. Starting from the touch dynamics, and keyboard dynamics nowadays, the main topic of study is mouse dynamics. Unlike the other techniques, mouse dynamics require simpler hardware to capture the biometric data without using sensitive data from the users. In this paper, we propose a method based on mouse dynamics and explainable deep learning for continuous and silent user authentication. We propose four different images obtained starting from mouse dynamics and we submit these images to several deep learning models obtaining interesting results in user behavior detection. Moreover, we propose the adoption of the Gradient-weighted Class Activation Mapping to highlight the areas of the images under study that is responsible for a specific categorization to explain the model decision.

Nowadays pathologists have to analyse blood cells manually with the aim to diagnose diseases. In order to perform this manual task blood samples must be collected from the patient and then placed on a microscope slide. This slide is studied with the aim to detect the abnormality presence. To automatise this process by helping pathologists, in this paper we propose the adoption of deep learning to automatically count and localise red blood cells, white blood cells and platelets by analysing blood microscopic images. We resort to the YOLO object detection model, able to look at the whole image so its predictions are informed by global context in the image. To show the effectiveness of the proposed method we evaluate our model on a dataset composed by 874 microscopic blood images, obtaining interesting results. Furthermore we show several examples related to how the proposed method can be helpful for the pathologists in their real-world work.

Diabetic retinopathy is a disease afflicting the retina and currently is manually diagnosed by specialists through eye tomography inspection. In order to assist the clinician in this time-consuming task, in this paper, we propose a method aimed to automatically diagnose the (proliferative and non-proliferative) diabetic retinopathy by exploiting deep learning. Furthermore, we investigate the possibility to automatically localise the areas related to the disease by exploiting class activation maps. We evaluate different deep learning models from a quantitative point of view (i.e, using metrics like accuracy, precision and recall) and a qualitative point of view (by exploiting class activation maps and image similarity metrics) with the aim to understand the quality of predictions performed by a model in retinopathy diagnosis, reducing the amount of knowledge required to assess the model performance. From the experimental analysis is emerging that deep learning shows an interesting diagnostic potential in the retinopathy disease localisation and can effectively help the clinician in retinopathy diagnosis. Moreover, the adoption of the class activation maps and its comparison evaluation can help the developers to debug the training step of the model without medical expertise.

Shallow machine learning is massively applied by researchers with the aim to detect (novel and unseen) malicious applications. Machine learning models are typically evaluated using malicious and trusted applications generated over a short period. In the real world, these models aim to identify malware that were not seen previously during the training phase. In this paper, we investigate how well machine learning-based malware detectors can actually detect malware in the real-world environment. By representing an Android application in terms of image, we evaluate the resilience of several popular supervised machine learning algorithms exploited by current literature for the malware detection task. The experimental results demonstrate the poor resilience of the machine learning models used for malware detection.

As a result of the spread in the mobile market, new kinds of malware have developed. Many malicious users began to produce harmful applications for open-source operating systems such as Android, and closed-source operating systems like iOS. For this reason, new anti-malware methodologies are necessary to identify them. Currently, most of them base their approach on the signature, and it does not allow users to defend themselves from current threats. In this article, we propose an automatic tool able to identify dangerous iOS applications by defining a system model through Milner's Calculus of Communicating Systems and the consequence use of the State Transitions System to evaluate the behaviors of an application, to categorize if it is malware or a legitimate application. As a result of the experiments conducted, we obtained relevant performances with good precision and recall levels.

Image segmentation is becoming a component of medical image processing with a significant role to analyze gross anatomy, to locate an infirmity and to plan the surgical procedures. Segmentation of brain magnetic resonance Imaging (MRI) is of increasing importance for the accurate diagnosis. For this reason, precise and accurate segmentation of brain MRI is a challenging task. In this paper we propose an approach for brain segmentation from MRI. Our method relies on a modified version of the U-Net convolutional neural network. Experiments performed on high grade brain cancer MRI demonstrate the effectiveness of the proposed approach for high grade brain cancer segmentation.

Unmanned aerial vehicles (drones) are widely available on today's market for use for use both as a hobby and to meet specific industrial needs, such as agriculture or parcel transport. This intensive use in various areas has also led to an increase in the illegitimate and criminal use of them, which has led to an interest in identification and detection techniques in the academic world. To address this topic, we propose a method that automates drone detection processes using their acoustic characteristics with different machine learning algorithms. In this study, we show the advantage of using machine learning techniques for drone detection. Experimental analysis demonstrates that the proposed method is promising in drone detection.

The advent and development of the smart grid have introduced a number of opportunities for improving efficiencies and overall performance. Along with the opportunities, each technology wave brings with it new opportunities for attackers to perpetrate damage, that in the smart grid scenario afflict electricity networks i.e. a cyber physical environment. In this paper we propose a method aimed to detect intrusion in smart grid by exploiting supervised machine learning techniques. By real-time log analysis, obtained from SCADA systems, widely exploited in a plethora of industrial applications, we extract a set of 26 features to build seven different machine learning models. Experimental analysis performed with 7 machine learning algorithms shows the effectiveness of the proposed method.

Laryngeal cancer tumor (LCT) grading is a challenging task in P63 Immunohistochemical (IHC) histopathology images due to small differences between LCT levels in pathology images, the lack of precision in lesion regions of interest (LROIs) and the paucity of LCT pathology image samples. The key to solving the LCT grading problem is to transfer knowledge from other images and to identify more accurate LROIs, but the following problems occur: 1) transferring knowledge without a priori experience often causes negative transfer and creates a heavy workload due to the abundance of image types, and 2) convolutional neural networks (CNNs) constructing deep models by stacking cannot sufficiently identify LROIs, often deviate significantly from the LROIs focused on by experienced pathologists, and are prone to providing misleading second opinions. So we propose a novel fusion attention block network (FABNet) to address these problems. First, we propose a model transfer method based on clinical a priori experience and sample analysis (CPESA) that analyzes the transfer ability by integrating clinical a priori experience using indicators such as the relationship between the cancer onset location and morphology and the texture and staining degree of cell nuclei in histopathology images; our method further validates these indicators by the probability distribution of cancer image samples. Then, we propose a fusion attention block (FAB) structure, which can both provide an advanced non-uniform sparse representation of images and extract spatial relationship information between nuclei; consequently, the LROI can be more accurate and more relevant to pathologists. We conducted extensive experiments, compared with the best Baseline model, the classification accuracy is improved 25%, and It is demonstrated that FABNet performs better on different cancer pathology image datasets and outperforms other state of the art (SOTA) models.

Recently, several research efforts have been focused on automotive safety, due to the increasing technology embedded in our vehicles. Research community have produced different methods aimed, for instance, to profile driver behaviour, starting from a feature set gathered by the vehicle. The provided methods are mainly machine learning-based: these solutions, as largely demonstrate in literature, suffer from several issues, due to the context variability but also because they are not able to provide a rational reason for the specific prediction. To overcome these limitations, in this paper we propose a novel model checking based approach to driver identification. Furthermore, a novel automatic procedure able to infer a logical representation of the driver behaviour is discussed. Two real-world datasets for the evaluation of the proposed method are considered, obtaining interesting results in driver identification.

Several techniques to overcome the weaknesses of the current signature based detection approaches adopted by free and commercial antimalware have been proposed by industrial and research communities. These techniques are mainly supervised machine learning based, requiring optimal class balance to generate good predictive models. In this paper, we propose a method to infer mobile application maliciousness by detecting the belonging family, exploiting formal equivalence checking. We introduce a set of heuristics to reduce the number of mobile application comparisons and we define a metric reflecting the application maliciousness. Real-world experiments on 35 Android malware families (ranging from 2010 to 2018) confirm the effectiveness of the proposed method in mobile malware detection and family identification.

Mobile malware is growing in number and its complexity is constantly increasing. Malware authors are continuously looking new ways to elude anti-malware controls. Anti-malware are not able to detect zero-day malware, because to detect malicious behaviour they need to know its signature, but to have this information the malware must already be widespread. Furthermore, anti-malware are able to scan one application at a time: for this reason a type of malware characterized by the colluding attack, where the malicious action is split in two (or more) applications, can not be recognised.

Thermal ablation refers to the destruction of tissue by extreme hyperthermia (elevated tissue temperatures) or hypothermia (depressed tissue temperatures). Radiofrequency ablation and microwave ablation are treatments that use image guidance to place a needle through the skin into a liver tumor. In radiofrequency ablation, high-frequency electrical currents are passed through an electrode in the needle, creating a small region of heat. In microwave ablation, microwaves are created from the needle to create a small region of heat. The heat destroys the liver cancer cells. Both the thermal ablation techniques are effective treatment options for patients who might have difficulty with surgery or those whose tumors are less than one and a half inches in diameter. In this paper, with the aim to help doctors in choosing of the best suitable thermal ablation treatment (i.e. with radiofrequency of micro waves), we propose a method exploiting machine learning to automatically predict the thermal ablation treatment starting from a set of features obtained from patients medical analysis.

Cybercriminals are continually working to develop increasingly aggressive malicious code to steal sensitive and private information from mobile devices. Antimalware are not always able to detect all threats, especially when they do not have previous knowledge of the malware signature. Moreover, malware code analysis remains a time-consuming process for security analysts. In this regard, we propose a method aimed to detect the malware belonging family and automatically pointing out a subset of potentially malicious classes. The rationale behind this work aims (i) to save valuable time for the security analyst by decreasing the amount of code to analyse, and (ii) to improve the interpretability of image-based deep learning model for malware family detection. We represent an application as an image and classify it with a deep learning model aimed to predict the belonging family; then, exploiting the use of activation maps, the approach points out potentially malicious classes to help the security analysts in the malicious behaviour recognition. The proposed method obtains an overall accuracy of 0.944 in the evaluation of a dataset composed of 8430 real-world Android malware, showing also that the use of activation maps can provide explainability about the deep learning model decision.

Social distancing is becoming really important in last month as a vehicle to limit the COVID-19 Coronavirus pandemic. Generally speaking, it is effective to control the spread of contagious diseases. In this context there is the need to monitor social distancing violations: for this reason in this paper we propose a social distancing detector able to count the violations by analysing video streams. Preliminary results show that the proposed method can be employed to guarantee social distancing. Moreover we discuss several suggestions aimed to improve the following proposal.

In this paper, we evaluate the performance of machine learning classifiers (Logistic Regression, CART, Random Forest) by fabricating adversarial examples (malware samples) statistically identical to goodware. To this end, we demonstrate three scenarios, (a) random attribute injection (b) insertion of prominent attributes from legitimate apps and (c) poisoning of class labels, for creating tainted malware samples, to mislead reduce accuracy of classification models. Experiments were conducted on data-set consisting of 15649 android applications comprising 5373 malicious and 10276 legitimate apps. The outcome of investigations demonstrates significant drop in accuracies in the range of 12-50%. However, in the absence of adversarial examples in the test set, the performance of classifiers was observed between 94.8-97.9%.

Smartphones, tablets and other mobile devices have become objects that we can no longer do without, as a matter of fact for us they are like an extension of our body and many people are addicted to them; this behavior is a consequence of the use we make of it, since these devices allow us to manage sensitive data (i.e., financial ones) and access information of different types (i.e., photos, messages or health data). For this reason it is essential to detect the harmful behaviors present within our smartphones, taking into account the weaknesses of the current anti-malware mechanisms. In this article we propose an approach capable of discriminating trusted applications from those that instead have malicious behavior, since they are involved in a colluding attack. We resort to the processing of the audio signal extracted from the conversion of an application into an audio file. The processing allows to generate a vector of characteristics to be analyzed with different classifiers. The experimental analysis is performed on a set of Android applications consisting of 359 trusted and (colluding) untrusted applications, showing the effectiveness of our method in detecting colluding applications.

Recently in the cybersecurity landscape, various figures have spread with different peculiarities. For instance there are the Black Hat hackers, aimed to perpetrate damage on the system or to silently exfiltrate sensitive information but there also the Ethical or White Hat hackers, aimed to investigate the vulnerabilities of a system under analysis only with the owner consent. In this context the Red Hat hackers, defined as vigilantes of the hacker world, are emerging. Their main aim is to independently found and solve vulnerabilities, by preventing cyberattacks. In this paper we propose a method aimed to automatise the vulnerability discover and mitigation process typically performed by Red Hat hackers. We exploit a tool-chain of several well-known tools and we evaluate the proposed method by exploiting the Metaesploitable Linux distro, showing that the proposed method is able to automatically mitigate vulnerabilities afflicting six widespread services..

The proliferation of info-entertainment systems in today’s vehicles has provided a really cheap and easy-to-deploy platform with the ability to gather information about the vehicle under analysis. Ultra-response connectivity networks with a latency below 10 milliseconds are providing the perfect infrastructure in which this information can be sent to improve safety and security. With the purpose of providing an architecture to increase safety and security in an automotive context, we in this paper propose a method for detecting the driver in real-time exploiting supervised machine learning techniques. The experimental analysis performed on real-world data shows that the proposed method obtains encouraging results.

The detection mechanism provided by current antimalware is the so-called signature based, requiring that a threat must be widespread to be recognised by the antimalware. Even if a malware is rightly recognized, by applying even trivial obfuscation techniques, it is really easy to bypass the antimalware detection mechanism. In this paper we propose a method to detect if an Android application is obfuscated with the call indirection obfuscation techniques by exploiting formal equivalence checking. In the experimental analysis we show the effectiveness of the propose approach for call indirection obfuscation technique detection, by exploiting two obfuscation tools.

OBJECTIVE: Due to the COVID-19 pandemic, our daily habits have suddenly changed. Gatherings are forbidden and, even when it is possible to leave the home for health or work reasons, it is necessary to wear a face mask to reduce the possibility of contagion. In this context, it is crucial to detect violations by people who do not wear a face mask. MATERIALS AND METHODS: For these reasons, in this article, we introduce a method aimed to automatically detect whether people are wearing a face mask. We design a transfer learning approach by exploiting the MobileNetV2 model to identify face mask violations in images/video streams. Moreover, the proposed approach is able to localize the area related to the face mask detection with relative probability. RESULTS: To asses the effectiveness of the proposed approach, we evaluate a dataset composed of 4095 images related to people wearing and not wearing face masks, obtaining an accuracy of 0.98 in face mask detection. DISCUSSION AND CONCLUSION: The experimental analysis shows that the proposed method can be successfully exploited for face mask violation detection. Moreover, we highlight that it is working also on device with limited computational capability and it is able to process in real time images and video streams, making our proposal applicable in the real world.

Modern cars include a huge number of sensors and actuators, which continuously exchange data and control commands. The most used protocol for communication of different components in automotive system is the Controller Area Network (CAN). According to CAN, components communicate by broadcasting messages on a bus. In addition, the standard definition of the protocol does not provide information for authentication, so exposing it to attacks. This paper proposes a method based on deep learning aiming at discovering attacks towards the CAN-bus. In particular, Neural Networks and MultiLayer Perceptrons are the class of networks employed in our approach. We also validate our approach by analysing a real-world dataset with the injection of messages from different types of attacks: denial of service, fuzzy pattern attacks, and attacks against specific components. The obtained results are encouraging and demonstrate the effectiveness of the approach.

Considering the pervasiveness of mobile devices, malicious writers are constantly focusing their attention in developing malicious payload aimed to gather sensible information from mobile devices without user content. As a matter of fact, it is really easy for malware writers to embed malicious payloads into legitimate applications, by applying the so-called repackaging paradigm, to generate a sample with a signature unknown to anti-malware software. In this paper we propose a twofold approach for the triage and the detection of repackaged Android applications. We propose a visualization schema to assist the malware analyst in the triage of unseen applications and a set of metrics for the automatic detection of repackaged applications. Experimental results show the effectiveness of the proposed approach.

Modern vehicles contain an ecosystem of several electronic units able to exchange data using the serial communication provided by the CAN bus. This protocol can be afflicted by a plethora of attacks that can expose the driver and the passengers to risks for their safety. In this paper we propose a method to detect potential attacks in automotive networks. We start from the analysis of a log obtained from a simulation and we consider a formal verification environment to verify whether the formal model we built from the log is safe. As a proof of concept, we evaluate the proposed method in a case study related to adaptive cruise control, to preliminarily demonstrate its effectiveness.

Medical image bio-markers of cancer are expected to improve patient care through advances in precision medicine. Compared to genomic bio-markers, bio-markers obtained directly from medical images provide the advantages of being a non-invasive procedure, and characterizing a heterogeneous tumor in its entirety, as opposed to limited tissue available for biopsy. In this paper, with the aim to demonstrate that noninvasive features can obtain better performances if compared to invasive ones in lung cancer detection and characterisation, we propose a method to discriminate between different lung cancers (i.e., Adenocarcinoma and Squamous Cell Carcinoma) by adopting both invasive (genomic) and non-invasive (radiomic) bio-markers, by building supervised machine learning models exploiting both invasive and non-invasive features. Experiments on a data-set of 130 patients show that radiomic bio-markers obtain better performances (with an f-measure equal to 0.993) if compared to the ones obtained by considering genomic ones (reaching an f-measure equal to 0.929) in lung cancer detection and characterisation.

Everyday born a new cyberattack and among these an emerging attack is represent by the so-called colluding. The application collusion attack is a new form of threat that is becoming widespread in mobile environment, especially in Android platform. This technique requires that two or more apps cooperate in some way with the aim to perform a malicious action that they are unable to perform independently. Detecting colluding apps is challenging problem, because currently there are no effective tools due to the search space of all possible combination of apps. In this paper we present a method exploiting model checking technique with the aim to detect a collusion attack between two applications. The method uses a heuristic function able to reduce the number of the analyzed apps and to localize the collusion attack. This heuristic function is based on the study of execution flow of an application, to identify the execution flow and verify it. The proposed algorithm verify if there is a flow of sensitive data that ends up in a shared resource and if this happens the app could be marked as potentially collusive, otherwise it is possible to exclude the app from the analysis, in order to reduce the number of apps to be analyzed. Experimental results on a data-set of Android applications show promising performances in colluding mobile app detection.

Artificial Intelligence systems have enabled significant benefits for users and society, but whilst the data for their feeding are always increasing, a side to privacy and security leaks is offered. The severe vulnerabilities to the right to privacy obliged governments to enact specific regulations to ensure privacy preservation in any kind of transaction involving sensitive information. In the case of digital and/or physical documents comprising sensitive information, the right to privacy can be preserved by data obfuscation procedures. The capability of recognizing sensitive information for obfuscation is typically entrusted to the experience of human experts, who are over-whelmed by the ever increasing amount of documents to process. Artificial intelligence could proficiently mitigate the effort of the human officers and speed up processes. Anyway, until enough knowledge won't be available in a machine readable format, automatic and effectively working systems can't be developed. In this work we propose a methodology for transferring and leveraging general knowledge across specific-domain tasks. We built, from scratch, specific-domain knowledge data sets, for training artificial intelligence models supporting human experts in privacy preserving tasks. We exploited a mixture of natural language processing techniques applied to unlabeled domain-specific documents corpora for automatically obtain labeled documents, where sensitive information are recognized and tagged. We performed preliminary tests just over 10.000 documents from the healthcare and justice domains. Human experts supported us during the validation. Results we obtained, estimated in terms of precision, recall and F1-score metrics across these two domains, were promising and encouraged us to further investigations.

Most of death causes are related to cardiovascular disease. In fact, there are several anomalies afflicting the heart beat, for instance heart murmur or artefact. We propose a method for heart disease detection. By gathering a set of feature obtainable directly from cardiac sounds, we consider this feature vector as input for a deep neural network to discriminate whether a cardiac sound is belonging to an healthy or to a patient with a cardiac disease. The experiment we performed demonstrated the effectiveness of the proposed approach in real-world environment.

At the end of 2019, a new form of Coronavirus, called COVID-19, has widely spread in the world. To quickly screen patients with the aim to detect this new form of pulmonary disease, in this paper we propose a method aimed to automatically detect the COVID-19 disease by analysing medical images. We exploit supervised machine learning techniques building a model considering a data-set freely available for research purposes of 85 chest X-rays. The experiment shows the effectiveness of the proposed method in the discrimination between the COVID-19 disease and other pulmonary diseases.

In last years vehicular network safety and security are attracting interest from both industries and researchers. In this paper, starting from a set of features gathered from the in-vehicle CAN bus, we show how machine learning algorithms can be useful to discriminate between the car owner and impostors. Furthermore, we assess several machine learning classifiers ability to predict instances not evaluated in the training set.

This paper presents a methodology for real-time data extraction and verification. In particular, considering the lacking of real-time data in sport analytics context, we propose a method to generate a data-set of player positions from soccer game videos, considering deep learning techniques, in order to extract player position, in terms of x-axis and y-axis, related to the accuracy of the detection. The experiment is performed on several local soccer games, captured using a stationary camera situated on the tribune area close to the center of the field. The camera is positioned in order to cover the entire soccer field. Clearly, this methodology can allow us to extract other types of information, such as distance between players and ball, and the area covered by all the players in a specific situation of the game. All of the informations are stored in a CSV (comma separated value) file that can be used to verify behavioural properties exploiting formal methods.

The application collusion attack is a new form of threat that is becoming widespread in mobile environment. This technique requires that two or more apps cooperate in some way with the aim to perform a malicious action that they are unable to perform independently. In this paper we propose the adoption of model checking to detect whether two or more apps are performing a collusion attack.

The current authentication systems based on password and pin code are not enough to guarantee attacks from malicious users. For this reason, in the last years, several studies are proposed with the aim to identify the users basing on their typing dynamics. In this paper, we propose the adoption of formal methods to discriminate between different users by exploiting a set of keystroke features. The idea behind the proposed method is to identify the users silently and continuously during their typing on a monitored system. To perform such user identification effectively, we consider a feature vector able to capture the typing style that is specific to each given user. By considering this feature model, in detail we propose to consider model checking with logic temporal properties to discriminate between different users using a set of keystroke features.

Data usually evolve according to specific processes, with the consequent possibility to identify a profile of evolution: the values it may assume, the frequencies at which it changes, the temporal variation in relation to other data, or other constraints that are directly connected to the reference domain. A violation of these conditions could be the signal of different menaces that threat the system, as well as: attempts of a tampering or a cyber attack, a failure in the system operation, a bug in the applications which manage the life cycle of data. To detect such violations is not straightforward as processes could be unknown or hard to extract. In this paper we propose an approach to model the data life cycle by observing the data evolution in its life cycle. Thus, we represent users able to alter data through timed automata. Through model checking, the obtained profile of evolution can be used to detect anomalies in relational database, data warehouse and big data.

Sport result prediction proposes an interesting challenge considering as popular and widespread are sport games, for instance tennis and soccer. The outcome prediction is a difficult task because there are a lot of factors that can afflict the final results and most of them are related to the player human behaviour. In this paper we propose a new feature set (related to the match and to players) aimed to model a soccer match. The set is related to characteristics obtainable not only at the end of the match, but also when the match is in progress. We consider machine learning techniques to predict the results of the match and the number of goals, evaluating a dataset of real-world data obtained from the Italian Serie A league in the 2017-2018 season. Using the RandomForest algorithm we obtain a precision of 0.857 and a recall of 0.750 in won match prediction, while for the goal prediction we obtain a precision of 0.879 in the number of goal prediction less than two, and a precision of 0.8 in the number of goal prediction equal or greater to two.

Currently, the gender is mainly inferred through bone and dental analyses. However, such sort of analysis is useless when bones are not available. Hair is a stable substance that, depending on its length, is capable of retaining years of information. In this paper we propose a machine learning based approach aimed to identify gender through hair elements. Preliminary results have even indicated that the method is promising also for forensics analysis since it is able to identify the gender with a high accuracy.

Lung cancer is one of primary premature death causes. As a matter of fact, people die more for lung cancer than prostate, colon or breast. To assist medics and radiologists in the diagnosis formulation, in this paper we propose a neural network-based method aimed to discriminate between different lung cancer types. We exploit a set of 30 radiomic feature directly obtained from magnetic resonance, tuning the neural network model when the momentum and the loss functions are varying with the aim to find the best model in terms of features and network parameters. We evaluate the effectiveness of the proposed method on a dataset of 2000 MRI labelled through medical reports, obtaining a precision equal to 0.918 and a recall equal to 0.923 in T1a lung cancer detection while a precision equal to 0.931 and a recall equal to 0.918 is obtained in T2b lung cancer detection.

Considering the increasing diffusion of mobile devices, attackers started to explore the possibility to perpetrate attacks using mobile surfaces (i.e., smartphones and tablets). Unfortunately, common antimalware techniques are often ineffective to detect new threats with the current signature based approach mainly adopted. In this we paper propose a set of features with the aim to discriminate between malware and trusted mobile applications: in detail we design a cascade learner where the first classifier of the cascade performs a coarse-grain analysis (it discriminates between malware and legitimate apps), while the second one performs a fine-grain analysis (it is aimed to identify the malware family). We obtain a precision equal to 0.947 and a recall equal to 0.962 in legitimate samples identification, while a precision equal to 0.961 and a recall equal to 0.946 is obtained in malware detection. With regard to family identification, an average precision and recall of 0.961 is obtained across 12 malware families.

Nowadays social networks are widespread used not only to enable users to share comments with other users but also as tool from which is possible to extract knowledge. As a matter of fact, social networks are increasingly considered to understand the opinion trend about a politician or related to a certain event that occurred: in general social networks have been proved useful to understand the public opinion from both governments and companies. In addition, also from the end users point of view it is difficult to identify real contents. This is the reason why in last years we are witnessing a growing interest in tools for analyzing big data gathered from social networks in order to find common opinions. In this context, content polluters on social networks make the opinion mining process difficult to browse valuable contents. In this paper we propose a method aimed to discriminate between pollute and real information from a semantic point of view. We exploit a combination of word embedding and deep learning techniques to categorize semantic similarities between (pollute and real) linguistic sentences. We experiment the proposed method on a dataset composed of real-world sentences gathered from the Twitter social network obtaining interesting results in terms of precision and recall.

Android malware is increasing from the point of view of the complexity and the harmful actions. As a matter fact, malware writers are developing sophisticated techniques to infect mobile devices very closed to their counterpart for personal computers. One of these threats is represented by the possibility to control the infected devices from the attacker i.e., the so-called botnet. In this paper a method able to identify botnet in Android environment through model checking is proposed. Starting from the malicious payload definition, the proposed method is able to detect and to localize the code related to the malicious botnet. We experiment real-world botnet based Android malware, obtaining encouraging results.

Blockchain technology is one of the most important and disruptive technologies in the world. Multiple industries are adopting the blockchain technology to innovate the way they work. One of the industries that are looking to adopt the blockchain is the healthcare industry. In fact, the protection of the private information stored in hospital database is a critical issue. In this paper we propose a method aimed to protect information exchanged in hospital networks, with particular regard to magnetic resonance images. As required from blockchain technology, each host network must validate the transiting data network: we exploit formal equivalence checking to perform this validation, by modeling magnetic resonance images in terms of automata by exploiting radiomic features.

Digital medical images are very easy to be modified for illegal purposes. An attacker may perform this act in order to stop a political candidate, sabotage research, commit insurance fraud, perform an act of terrorism, or even commit murder. Between the machine that performs medical scans and the radiologist monitor, medical images pass through different devices: in this chain an attacker can perform its malicious action. In this paper we propose a method aimed to avoid medical images modifications by means of equivalence checking. Magnetic images are represented as finite state automata and equivalence checking is exploited to check whether the medical resource have been subject to illegal modifications.

The picture archiving and communication system is a medical imaging technology used primarily in healthcare organizations to store and digitally transmit electronic images and clinically-relevant reports. As demonstrated, these systems can be exploited by malicious users: in fact, considering that medical images are not digitally encrypted, any medical image modifications would be difficult to detect for a radiologist. To mitigate this aspect, in this paper a formal modelisation for picture archiving and communication system systems is proposed. The main aim is to avoid illegal writing and reading from components that should not do it, by representing the system components in terms of automa.

Formal verification allows checking the design and the behaviour of a system. One of the main limitations to the adoption of formal verification techniques is the process of model creation using specification languages. For this reason a tool supporting this activity is necessary. Actually, there are several tools allowing analysts to verify models expressed into specification languages. These tools provide support for automatically checking whether a system satisfies a property. However, to use such tools it is important to deeply know a precise notation for defining a system, i.e., the Calculus of Communicating Systems. Since systems are often expressed as time-series, to overcome this problem, we provide an user-friendly tool able to automatically generate a system model starting from the CSV - Comma-Separated Values format (the most widespread format considered to release dataset). In this way we hide the details about the model construction form the analyst, which can only focus immediately on the properties to verify. We introduce Dunuen , a tool allowing the user to firstly perform a kind of pre-processing operation starting from a CSV file, as discretization or removing attributes; subsequently it automatically creates a formal model from the pre-processed CSV file and, by invoking the model checker we embedded in Dunuen , it finally verifies whether the generated model satisfies a property expressed in temporal logic through a graphic interface, proposing formal methods as an alternative to machine learning for classification tasks.

Data tipically evolve according to specific processes, with the consequent possibility to identify a profile of evolution: the values it may assume, the frequencies at which it changes, the temporal variation in relation to other data, or other constraints that are directly connected to the reference domain. A violation of these conditions could be the signal of different menaces that threat the system, as well as: attempts of a tampering or a cyber attack, a failure in the system operation, a bug in the applications which manage the life cycle of data. To detect such violations is not straightforward as processes could be unknown or hard to extract. In this paper we propose an approach to detect data anomalies. We represent data user behaviours in terms of labelled transition systems and through the model checking techniques we demonstrate the proposed modeling can be exploited to successfully detect data anomalies.

The ineffectiveness of signature-based malware detection systems prevents the detection of malware, even objects of trivial obfuscation techniques, makes mobile devices vulnerable. In this paper a dynamic technique to detect malware on Android platform is proposed. We exploit a set of energy related features i.e., feature which can be symptomatic of abnormal battery consumption. We built different models exploiting four different supervised machine learning classification algorithms, obtaining for all the evaluated models an accuracy greater than 0.91.

Human activity recognition is attracting interest from researchers and developers in recent years due to its immense applications in wide area of human endeavors. The main issue in human behaviour modeling is represented by the diverse nature of human activities and the nature in which they are performed by the individual makes them challenging to recognize. In this paper we propose a method aimed to recognize human activities and detect users using features gathered from accelerometer sensors widespread in wearable and mobile devices. We exploit machine learning aimed to build models with the ability to discriminate between a set of user activities: sitting, sitting down, standing, standing up and walking. Furthermore, we demonstrate that the proposed method is able to distinguish between different users and to identify the user genre. Real-world experiment shows the effectiveness of the proposed solution.

The European Railway Traffic Management System has the purpose to provide a common signaling system for all the European nations. It consists of two subsystems: the trackside subsystem (TSS) and the on-board subsystem (OBS) that communicate to exchange information about the state of the trackside and/or the train. Radio communication can take place according to the requirements specification reported in ERTMS/ETCS SUBSET-026-3. As the communication between TSS and OBS is a critical issue, we exploit model checking to verify the correctness of the communication process as specified in the SUBSET-026-3. The results achieved during the experimentation seem to be very promising.

SCADA control systems use programmable logic controller to interface with critical machines. SCADA systems are used in critical infrastructures, for instance, to control smart grid, oil pipelines, water distribution and chemical manufacturing plants: an attacker taking control of a SCADA system could cause various damages, both to the infrastructure but also to people (for instance, adding chemical substances into a water distribution systems). In this paper we propose a method to detect attacks targeting SCADA systems. We exploit model checking, in detail we model logs from SCADA systems into a network of timed automata and, through timed temporal logic, we characterize the behaviour of a SCADA system under attack. Experiments performed on a SCADA water distribution system confirmed the effectiveness of the proposed method.

With reference to a distributed environment consisting of nodes connected in an arbitrary network topology, we propose the organization of a protection system in which a set of subjects, e.g. processes, generates access attempts to memory segments. One or more primary passwords are associated with each node. An access to a given segment can be successfully accomplished only if the subject attempting the access holds an access privilege, certified by possession of a valid protected pointer (p-pointer) referencing that segment. Each p-pointer includes a local password; the p-pointer is valid if the local password descends from a primary password by application of a universally known, parametric one-way generation function. A set of protection primitives makes it possible to manage the primary passwords, to reduce p-pointers to include less access rights, to allocate new segments, to delete existing segments, to read the segment contents and to overwrite these contents. The resulting protection environment is evaluated from a number of viewpoints, which include p-pointer forging and revocation, the network traffic generated by the execution of the protection primitives, the memory requirements for p-pointer storage, security, and the relation of our work to previous work. An indication of the flexibility of the p-pointer concept is given by applying p-pointers to the solution of a variety of protection problems.

With the increasing diffusion of mobile technologies, nowadays mobile devices represent an irreplaceable tool to perform several operations, from posting a status on a social network to transfer money between bank accounts. As a consequence, mobile devices store a huge amount of private and sensitive information and this is the reason why attackers are developing very sophisticated techniques to extort data and money from our devices. This paper presents the design and the implementation of LEILA (formaL tool for idEntifying mobIle maLicious behAviour), a tool targeted at Android malware families detection. LEILA is based on a novel approach that exploits model checking to analyse and verify the Java Bytecode that is produced when the source code is compiled. After a thorough description of the method used for Android malware families detection, we report the experiments we have conducted using LEILA. The experiments demonstrated that the tool is effective in detecting malicious behaviour and, especially, in localizing the payload within the code: we evaluated real-world malware belonging to several widespread families obtaining an accuracy ranging between 0.97 and 1.

Sport analytics is the process of searching, interpreting and processing information related to the sports performance of a team or a single athlete, using information systems and mathematical methods of data evaluation. The aim is to encourage improvement of sports performance. In this paper we propose a formal methodology aimed to design an expert system to support real-time decisions for notational analysis in sport environment exploiting big data obtained from several sources. The aim is to improve and predict individual and collective performances during a match.

Considering the widespread diffusion of mobile applications in many aspects of our life, one of the most required developer's skill is related to mobile applications development, with particular regard to the Android environment. In this paper, we present a framework aimed to quantify students learning ability in mobile applications development. The proposed framework can be useful also to teachers in order to improve their teaching method.

Despite the development of new technologies, in order to prevent the stealing of cars, the number of car thefts is sharply increasing.With the advent of electronics, new ways to steal cars were found.To avoid auto-theft attacks, in this paper we propose a machine leaning based method to silently e continuously profile the driver by analyzing built-in vehicle sensors.We evaluate the efficiency of the proposed method in driver identification using 10 different drivers.Results are promising, as a matter of fact we obtain a high precision and a recall evaluating a dataset containing data extracted from real vehicle.

In the last years, several safety automotive concepts have been proposed, for instance the cruise control and the automatic brakes systems.The proposed systems are able to take the control of the vehicle when a dangerous situation is detected.Less effort was produced in driver aggressiveness in order to mitigate the dangerous situation.In this paper we propose an approach in order to identify the driver aggressiveness exploring the usage of unsupervised machine learning techniques.A real world case study is performed to evaluate the effectiveness of the proposed method.

Data breaches usually involve financial information such as credit card or bank details. Automated formal verification of safety critical systems has been mostly focused on analysing high-level abstract models which, however, are significantly different from real implementations written in programming languages. In this paper we propose a technique that links model checking verification closer to real implementations, with particular regard to financial environment, in order to identify possible causes of data breaches. A formal model starting from execution traces is retrieved. Thus, the discovered model can be analysed to verify it respects the defined properties. A real safety critical system has been used as a case study to evaluate the proposed methodology.

The crucial aim of software security is malware detection. A malware is a program with malicious intents. The predominate anti-malware solutions are signature-based. These detectors compute the signature starting from the syntactic characteristics of the malicious code. Unfortunately, the signature-based techniques are ineffective against the code obfuscations, i.e., trivial transformations that alter the syntax of the code preserving the normal behaviour of the program. To address this limitation, formal methods are used in software security. Formal methods are rigorous techniques used to verify the behaviour of a system. This paper aims to make an overview on behavioural based techniques developed to detect malware programs. The illustrated approaches are based on different formal techniques.

Android operating system is constantly overwhelmed by new sophisticated threats and new zero-day attacks. While aggressive malware, for instance malicious behaviors able to cipher data files or lock the GUI, are not worried to circumvention users by infection (that can try to disinfect the device), there exist malware with the aim to perform malicious actions stealthy, i.e., trying to not manifest their presence to the users. This kind of malware is less recognizable, because users are not aware of their presence. In this paper we propose FormalDroid, a tool able to detect silent malicious beaviours and to localize the malicious payload in Android application. Evaluating real-world malware samples we obtain an accuracy equal to 0.94.

Modern vehicles have lots of connectivity, this is the reason why protect in-vehicle network from cyber-attacks becomes an important issue. The Controller Area Network is a de facto standard for the in-vehicle network. However, lack of security features of CAN protocol makes vehicles vulnerable to attacks. The message injection attack is a representative attack type which injects fabricated messages to deceive original Electronic Control Units or to cause malfunctions. In this paper we propose a method able to detect four different type of attacks targeting the CAN protocol adopting fuzzy algorithms. We obtain encouraging results with a precision ranging from 0.85 to 1 using the fuzzy NN algorithm in the identification of attacks targeting CAN protocol.

Medical studies demonstrated that diabetes pathology is increasing in last decades and the trend do not tends to stop. In order to help and to accelerate the diagnosis of diabetes in this paper we propose a method able to classify patients affected by diabetes using a set of characteristic selected in according to World Health Organization criteria. Evaluating real-world data using state of the art machine learning algorithms, we obtain a precision value equal to 0.770 and a recall equal to 0.775 using the HoeffdingTree algorithm.

Android represents the most widespread mobile environment. This increasing diffusion is the reason why attackers are attracted to develop malware targeting this platform. Malware writers usually use code obfuscation techniques in order to evade the current antimalware detection and to generate new malware variants. These techniques make code programs harder to understand and they change the signature of the application making ineffective the signature extraction work. We propose a method based on formal methods able to identify whether a mobile application is obfuscated. In this preliminary work we identify one of the most widespread obfuscation technique: the code reordering. We test our method on a real-world dataset composed by Android trusted and ransomware samples, obtaining encouraging results.

Android is currently the most widely used mobile environment. This trend encourages malware writers to develop specific attacks targeting this platform with threats designed to covertly collect data or financially extort victims, the so-called ransomware. In this paper we use formal methods, in particular model checking, to automatically dissect ransomware samples. Starting from manual inspection of few samples, we define a set of rule in order to check whether the behaviours we find are representative of ransomware functionalities.

Android smartphones are the most widespread in the world. This is the reason why attackers write code more and more aggressive in order to steal data and other important information stored in the phone. One of the most representative malware that implements the typical trojan behaviour in Android environment is the so-called Fake Installer. In this paper we use formal methods, in particular model checking, in order to identify Fake Installer malware. We specify a set of formulae and then we check these on a designed application model, built in CCS, to recognize whether an application is a malware belonging to Fake Installer family or a legitimate sample. We experiment our methodology on 1125 real world samples obtaining very promising results.

In this paper, we use model checking to perform the analysis and the assessment of the Purchasing Process performance in the field of Action Controls. The aim of these controls, in a Management Control Systems, is to ensure that the employee behaviours are consistent with the objectives and strategies of the organization. However, model checking suffers from the so-called state explosion problem, which says that the state space grows exponentially in the number of concurrent processes. In this paper we consider a property-based methodology developed to combat the state explosion problem. Our focus is two fold: (i) we show how model checking can be applied in the context of business process modelling and analysis, (ii) we evaluate and test the reduction methodology using an academic case study. Our investigations suggest that the business community can benefit from this efficient methodology developed in formal methods since it can detect errors that were missed by traditional verification techniques. Thus, formal methods can be adopted to enhance the effectiveness of the organizational control together with an increase of process efficiency levels. We show and discuss the obtained experimental results.

A wiki is a collaborative Web site whose content can be edited by anyone who has the access. Wikis are becoming a new work tool in enterprises and are widely spreading every-where. Indeed, it is important to consider the design and evolution of a wiki. We present a methodology to help wiki designers, engineers and domain experts. In practice, from the wikis we derive formal models, which are successively used to perform equivalence checking. More precisely, in order to design a wiki p we propose a methodology for the selection of the best-fit wiki model q, among a set of candidate ones. For best fit we mean that p and q have a similar structure. To handle the complexity of finding all possible candidates processes q, a heuristic function can be used to filter the set of significant candidates, and to speed up the search of a successful one, which is the main contribution of the paper. Eventually, the elected model may be exploited to start the design process.

Wikis delineate a new work tool in enterprises and they are spreading everywhere. Indeed, they are often used as internal documentation for various in-house systems and applications as well as powerful tools for collaboration and knowledge sharing. As occurs with software, the fundamental growth of a wiki may lead to its degradation. The quality of wikis, especially in enterprise contexts, should not play a trivial role. Software quality is a very discussed topic, but there are not many studies regarding the quality of wikis. We propose a probabilistic model to represent wikis and to investigate their quality. Due to the similarity with the World Wide Web it is natural to consider the popular Google PageRank (with minor modifications) to calculate probabilities between pages. Each wiki category, a set of wiki pages, is modelled using the PRISM language in order to verify specific properties in PCTL*. Experiments conducted on a adequate number of (enterprise) wikis assess the validity of our methodology.

Equivalence checking is a powerful formal technique to improve the quality of computer and software systems. It is usually employed to verify the correctness in a model-based design. Notwithstanding, a detailed and precise specification is required in order to apply equivalence checking to a given domain. Unfortunately, certain fields of application like business process management lack of such adequate information. We explore the applicability of equivalence checking to validation of Business Processes described by the aid of Workflow Management systems. Due to the state explosion problem, formal methods are not very popular in the business domain. In fact, the state space grows exponentially in the number of concurrent processes leading to an impracticable verification. In this paper we deal with a heuristic-based methodology developed to beat the state explosion problem when checking non-equivalence. Our contribution is two-fold: (i) we show how equivalence checking can successfully operate in the business modelling and analysis context, (ii) we model and use the bank supply process as a real case study to evaluate and test the heuristic-based methodology. We show and debate encouraging experimental results comparing them with a state of the art model checker i.e. CADP. This suggests that the business community, mostly in the banking field, can take advantage from our efficient methodology based on process algebra.

We present ELOTON, an Eclipse-based Editor to help people who want to approach the Language of Temporal Ordering Specification (LOTOS). LOTOS is a Formal Description Technique standardized by ISO for the design of concurrent and distributed systems, and in particular for OSI services and protocols. LOTOS has been widely used for the specification of large data communication systems. It is mathematically well-defined and expressive: it allows concurrency, non-determinism, synchronous and asynchronous communications. CADP is a popular toolbox that supports high-level descriptions written in LOTOS, among others. Unluckily, there not exists an user interface suitable for newcomers in formal methods. Thus, many people encounter many obstacles in using formal methods and in particular model checking due to the lacking of user-friendly tools. We argue that ELOTON, thanks to its rich text editor and visualization features might involve a major number of users coming from various disciplines. Our tool serves as a graphical front-end for CADP.

A wiki is a collaborative Web site whose content can be edited by anyone who has access to it. Wikis are becoming a new work tool in enterprises and are widely spreading everywhere. In fact, they are often used as internal documentation for various in-house systems and applications. Understanding and maintaining the structure of a wiki may be a crucial aspect. As well as software grows, decays and needs refactoring, the organic growth of a wiki inevitably leads to its degradation. We propose a novel methodology based on formal methods to analyse and verify the architecture of wikis. Formal verification helps to perform refactoring. Each wiki category, a set of wiki pages, is modelled using the Calculus of Communicating Systems (CCS) process algebra in order to verify specific properties. First experiments conducted on a adequate number of categories of Wikipedia, assess the validity of our methodology revealing new directions for future research.

Equivalence checking plays a crucial role in formal verification to ensure the correctness of concurrent systems. However, this method cannot be scaled as easily with the increasing complexity of systems due to the state explosion problem. This article presents an efficient procedure, based on heuristic search, for checking Milner's strong and weak equivalence; to achieve higher efficiency, we actually search for a difference between two processes to be discovered as soon as possible, thus the heuristics aims to find a counterexample, even if not the minimum one, to prove nonequivalence. The presented algorithm builds the system state graph on-the-fly, during the checking, and the heuristics promotes the construction of the more promising subgraph. The heuristic function is syntax based, but the approach can be applied to different specification languages such as CCS, LOTOS, and CSP, provided that the language semantics is based on the concept of transition. The algorithm to explore the search space of the problem is based on a greedy technique; GreASE (Greedy Algorithm for System Equivalence), the tool supporting the approach, is used to evaluate the achieved reduction of both state-space size and time with respect to other verification environments.

Reverse engineering of regulatory relationships from genomics data is emerging as crucial to dissect the complex underlying regulatory mechanism occurring in a cell. In this paper we propose a novel reverse engineering algorithm that makes use of formal methods, usually adopted in engineering to specify and verify concurrent software and hardware systems. With a formal specification of gene regulatory hypotheses we are able to prove mathematically whether a time course experiment belongs or not to the formal specification, determining in fact whether a gene regulation exists or not. The method is capable to detect both direction and sign (inhibition/activation) of regulations whereas most of literature methods which are limited to undirected and/or unsigned relationships. The method, empirically evaluated on experimental and synthetic datasets, reaches high levels of accuracy, outperforming literature methods in terms of precision and recall, despite the computational cost increases exponentially with the size of the network.

Model checking is a very useful method to verify concurrent and distributed systems which is traditionally applied to computer system design. We examine the applicability of model checking to validation of Business Processes that are mapped through the systems of Workflow Management. The use of model checking in business domain is affected by the state explosion problem, which says that the state space grows exponentially in the number of concurrent processes. In this paper we consider a property-based methodology developed to combat the state explosion problem. Our focus is two fold; firstly we show how model checking can be applied in the context of business modelling and analysis and secondly we evaluate and test the methodology using as a case study a real-world banking workflow of a loan origination process. Our investigations suggest that the business community, especially in the banking field, can benefit from this efficient methodology developed in formal methods since it can detect errors that were missed by traditional verification techniques, and being cost-efficient, it can be adopted as a standard quality assurance procedure. We show and discuss the experimental results obtained.

Ensuring deadlock freedom is one of the most critical requirements in the design and validation of concurrent systems. The biggest challenge toward the development of effective deadlock detection schemes remains the state-space explosion problem when model checking is used for proving the correctness of a system with respect to a desired behavior. In this paper we propose the use of the Ant Colony Optimization (ACO) to reduce the state explosion problem arising when finding deadlocks in complex networks described using Calculus of Communicating Systems (CCS). Moreover, ACO is used to provide minimal counterexamples. In fact, although one of the strongest advantages of model checking is the generation of counterexamples when verification fails, traditional model checkers may return very long counterexamples. We present an implementation of our technique and encouraging experimental results on several benchmarks. These results are then compared with other heuristic-based search strategies, retaining the advantages of our approach.

Abstract In this paper, we propose a bottom‐up approach for the verification of systems with modular structure: we prove that when the modules are composed in specific ways, the complete software system verifies a composition of the properties each component does. We focus on the process of upgrading systems with new functionalities, where the validity of old requirements needs to be ensured, but also an understanding of the new properties the upgraded system would enjoy is useful. In this work, we assume each component to be specified by a CCS process, and the properties to be expressed by selective mu‐calculus formulae. Copyright © 2007 John Wiley & Sons, Ltd.

Many tools for the automatic analysis or verification of finite-state distributed systems are based on construction of the global state graph of the system under consideration. Thus, they often fail because of the state explosion problem: the state space of a distributed system potentially increases exponentially in the number of its parallel components. To overcome this problem, we present a model checking procedure, based on the combination of heuristic searches with ideas taken from local model checking. We use heuristic mechanisms for exploration of the search space in order to avoid construction of the complete state graph.

When verifying applications that require a high level of reliability, testing fails to assure an adequate level of correctness. Thus we apply model checking, an automatic verification technique: the programs are written in a simple sequential language and properties are specified by a suitable temporal logic. We obtain a finite state representation of the system by applying abstraction techniques. The abstraction is derived from the logic formula representing the property to be checked. A tableau-based method is developed to prove satisfaction.

We present a model checking procedure to verify a subset of the Java Virtual Machine Language. The procedure works on a finite tractable state representation of the program: in fact, abstraction techniques are employed, driven by the temporal logic formula representing the property to be checked. A tableau-based method is developed to proof the satisfaction of the formulae: thus not all the program computations are checked, but only those ones interested in the goal of the property verification.

Abstract A well‐known problem in the verification of concurrent systems based on model checking is state explosion: concurrent systems are often represented by automata with a prohibitive number of states. A reduction technique to reduce state explosion in deadlock checking is presented. The method is based on an automatic syntactic simplification of a calculus of communicating systems (CCS) specification, which keeps the parts of the program structure that may lead to a deadlock and deletes the other parts. Copyright © 2002 John Wiley & Sons, Ltd.

Abstract A system supporting video on demand is modeled in the process calculus CCS (Calculus of Communicating Systems), while some properties are expressed in a temporal logic and verified by means of the model checkers of the North Carolina Concurrency Workbench. This application was chosen as a case study to evaluate the usefulness of a methodology, by means of which a property is checked on reduced models obtained issuing abstractions of the system on the basis of the formula. Experimental results are shown and discussed. Copyright © 2001 John Wiley & Sons, Ltd.

We present the formal verification of a multicast protocol for mobile computing. The protocol supports reliable and totally ordered communication within a set of processes running on mobile hosts. Mobile hosts communicate with a wired infrastructure through wireless links. The protocol is specified in Calculus of Communicating Systems and checked using the Concurrency Workbench tool. The protocol was chosen as a case study to evaluate the usefulness of a methodology, by means of which a property is checked on a reduced system, where the reduction is driven by the formula expressing the property itself. The reduction is obtained by transforming the program into one having a smaller representation. The approach is based on a logic, the selective mu-calculus, which has the characteristic that each formula allows the immediate pointing out of the parts of the system that do not alter the truth value of the formula itself, and thus can be ignored. We show and discuss the experimental results obtained.

LOTOS is a formal specification language for concurrent and distributed systems. Basic LOTOS is the version of LOTOS without value-passing. A widely used approach to the verification of temporal properties is model checking. Often, in this approach the formal specification is translated into a labeled transition system on which formulae expressing properties are checked. A problem with this verification technique is state explosion: concurrent systems are often represented by automata with a prohibitive number of states. In this paper we show how, given a set ρ of actions, it is possible to automatically obtain for a Basic LOTOS program a reduced transition system to which only the arcs labeled by actions in ρ belong. The set ρ of actions plays a fundamental role in conjunction with a temporal logic defined by the authors in a previous paper: selective mu-calculus. The reduced system with respect to ρ preserves the truth value of all selective mu-calculus formulae with actions from the set ρ. We act at both syntactic and semantic levels. From a syntactic point of view, we define a set of transformation rules obtaining a smaller program. On the semantic side, we define a non-standard semantics which dynamically reduces the transition system during generation. We present a tool implementing both the syntactic and the semantic reduction. Copyright © 1999 John Wiley & Sons, Ltd.