• Proposal of three architectures for biometric airport check-in systems. • Comparisonofcentralized and federated architecture to remark privacy preserving issues. • Quantitative and Qualitative Assessment for privacy analysis. • Trade-off analysis between privacy and accuracy in biometric systems. • Federated Learning-based strategies for privacy preservation. The rapid adoption of automated airport check-in systems using facial recognition raises significant privacy concerns due to their reliance on centralized deep learning models that store and transmit biometric data from edge devices. While Federated Learning (FL) is a promising approach for privacy preservation, its effectiveness in biometric identification remains underexplored, particularly in real-world environments like airports. This study assesses the privacy implications of FL in facial recognition by comparing three architectures. A first centralized system, where biometric data is sent to a central server for model training and inference, posing significant privacy risks. The second is a one-level FL architecture, where biometric data remains on local devices, and only model updates are shared with a central aggregator. The third is a two-level FL architecture, introducing an additional aggregation layer among airlines to enhance model generalization while preserving privacy. To ensure a rigorous privacy preservation evaluation, we integrate both quantitative and qualitative metrics. For the quantitative assessment, we leverage the Privacy Meter Tool, which enables simulations of Membership Inference Attacks and the application of Differential Privacy as a mitigation technique. For the qualitative evaluation, we conduct a Data Protection Impact Assessment to analyze potential privacy risks from a regulatory perspective. Additionally, we assess model accuracy, computational efficiency, and communication overhead to determine FL’s feasibility in large-scale airport environments. Our results show that while FL significantly reduces privacy risks, the two-level FL approach introduces new vulnerabilities, such as model poisoning risks and privacy-utility trade-offs, requiring further mitigation strategies like DP.

The growing interest in artificial intelligence applications in real clinical practice has made the development of personalised medicine possible. Digital Twins support diagnosis and treatment by providing an overall view of patients’ health status. The definition of a patient’s digital model requires the integration of different sources of information that contribute to a holistic view of the subject. In this work, we propose a preliminary step toward the definition of paediatric digital twins, providing a tool for Region of Interest identification on X-ray images. In detail, the proposed tool, STELLA (Segmentation Tool for Enhanced Localisation and Labelling of diagnostic Areas), is adopted to automatically detect the bladder and urethra regions on the images obtained from the cystourethrography exam. STELLA pipeline is based on Segment Anything Model (SAM) for the segmentation task and Resnet-18 for masks classification: SAM is leveraged for automatic masks generation and ResNet18 is trained on labelled masks for Regions of Interest classification. This is framed in a larger context, whose aim is to support posterior urethral valves diagnosis.

In this era of huge data availability, data-driven approaches are affirming themselves as one of the dominant paradigm in model identification. Due to their capability to fit acquired data, these models exhibit flexibility and the ability to cope with undiscovered knowledge. This paper proposes a method to overcome existing limitations in the model and parameter identification of complex stochastic Time-Series, enabling the identification of processes characterisable according to the Gaussian Mixture Model. More concretely, this paper aims to define methods for learning and classifying the model of complex stochastic processes.

This work explores the benefits of applying swarm optimization strategies to tackle cybersecurity challenges and to model both defense mechanisms and threat intelligence solutions—particularly those involving Generative AI, though not limited to them. Among the various swarm optimization algorithms, this study focuses on Cat Swarm Optimization (CSO). Specifically, CSO is employed to address a feature selection task, with the selected features subsequently used to train a Generative Adversarial Network (GAN) for cyber attack identification. CSO was chosen because the behavioral patterns of cats offer a compelling analogy for modeling the stealthy and dynamic nature of both cyberattacks and defensive responses. A case study, represented by a SQL-Injection attack, is also presented to provide a preliminary evaluation of the proposed approach.

Human Digital Twin is an emerging technology that could revolutionize the current healthcare system by enabling the delivery of Personalized Health Services through the use of tools such as Artificial intelligence. However, the considerable complexity of the structure of the human body, brought about by continuous molecular and physiological changes, makes it extremely difficult to process medical data extracted by Artificial intelligence techniques. The latter requires a large amount of data for reliable performance, which is often difficult to obtain due to limited quality and availability. In this paper, we propose a methodology to generate Artificial medical data. In detail, we focus on generating Artificial voice signals. The analysis of voice recordings is fundamental to diagnose specific pneumo-articulatory apparatus diseases, such as dysphonia. The generative neural network employed is based on the WaveNet model, due to its autoregressive sampling, which enables generating recordings of variable length. We propose a setup which enables to generate Artificial samples of required sex and pathology to balance and augment the dataset using only one generative network. The quality of the generative network is assessed by balancing the training dataset by generated data and training a convolutional classifier, which is tested on a dataset which was not introduced to the generative network during training. We achieved reasonable improvements in classification accuracy, particularly for the under-represented sex in terms of accuracy, arguing that this approach is worthy of future research.

Recently, Large Language Models (LLMs) have seen some impressive leaps, achieving the ability to accomplish several tasks, from text completion to powerful chatbots. The great variety of available LLMs and the fast pace of technological innovations in this field, is making LLM assessment a hard task to accomplish: understanding not only what such a kind of systems generate but also which is the quality of their results is of a paramount importance. Generally, the quality of a synthetically generated object could refer to the reliability of the content, to the lexical variety or coherence of the text. Regarding the quality of text generation, an aspect that up to now has not been adequately discussed is concerning the readability of textual artefacts. This work focuses on the latter aspect, proposing a set of experiments aiming to better understanding and evaluating the degree of readability of texts automatically generated by an LLM. The analysis is performed through an empirical study based on: considering a subset of five pre-trained LLMs; considering a pool of English text generation tasks, with increasing difficulty, assigned to each of the models; and, computing a set of the most popular readability indexes available from the computational linguistics literature. Readability indexes will be computed for each model to provide a first perspective of the readability of textual contents artificially generated can vary among different models and under different requirements of the users. The results obtained by evaluating and comparing different models provide interesting insights, especially into the responsible use of these tools by both beginners and not overly experienced practitioners.

Large Language Models (LLMs) are rapidly evolving, demonstrating impressive capabilities in multimedia objects generation, ranging from text and image generation from scratch to programming code and efficient conversational agents. From the perspective of cyber-security challenges, LLMs and cyber-security are in a controversial relationship: it can be observed that LLMs, as a type of AI, play a mainfold role: that of security guardians, that of security breachs ”unaware” enablers and that of victims of cyber attacks. In fact, LLMs are able to enhance security of several tasks and applications but they are also attractive for malicious users to be exploited as means to perform novel attacks and, finally they represent challenging assets for targeting attacks. In this work, we discuss this mainfold key reading by providing a brief landscape of both the current defence applications of LLMs against cyber attacks and the currently known LLMs security vulnerabilities along with potential cyber-attacks targeting and involving LLMs. The final aim of study is intended to provide a guideline to further explore specific cyber-security scenarios involving LLMs.

Evaluating the quality of synthetically generated data is an open problem, still not adequately debated when compared to the interest in how it could be generated, and still far from an effective standardization stage. The evaluation of such artificial data produced, for example, by GANs or other generative AI models, should be approached having in mind several relevant questions concerning, among the others, the following ones: i) the meaning of reliability, faithfulness, and overall quality of artificial data as respect to genuine data; ii) the appropriate set of metrics and tools to estimate the quality of synthetic data; iii) the identification of dependencies occurring among the quality of the artificial data, the generative models to produce them and the kind of native data used to train the generative processes. This work aims to contribute to this challenge by providing a reconnaissance study supporting the groundwork for a more structured analysis of the problem. More exactly, a particular kind of Generative models, such as the Tabular GANs, is discussed in details to focus the attention on effective methods for analyzing and of evaluating the quality of synthetically generated data.

Human-Machine Interfaces (HMIs) are getting more and more important in a hyper-connected society. Traditional HMIs are built considering cognitive features while emotional ones are often neglected, bringing sometimes such interfaces to misuse. As a part of a long run research, oriented to the definition of an HMI engineering approach, this paper concretely proposes a method to build an emotional-aware explicit model of the user starting from the behaviour of the human with a virtual agent. The paper also proposes an instance of this model inference process in voice assistants in an automatic depression context, which can constitute the core phase to realize a Human Digital Twin of a patient. The case study generated a model composed of Fluid Stochastic Petri Net sub-models, achieved after the data analysis by a Support Vector Machine.

In a world whose attention to environmental and health problems is very high, the issue of properly managing nuclear waste is of a primary importance. Information and Communication Technologies have the due to support the definition of the next-generation plants for temporary storage of such wasting materials. This paper investigates on the adoption of one of the most cutting-edge techniques in computer science and engineering, i.e. Digital Twins, with the combination of other modern methods and technologies as Internet of Things, model-based and data-driven approaches. The result is the definition of a methodology able to support the construction of risk-aware facilities for storing nuclear waste.

The diffusion of intelligent services and the push for the integration of computing systems and services in the environment in which they operate require a constant sensing activity and the acquisition of different information from the environment and the users. Health monitoring, domotics, Industry 4.0 and environmental challenges leverage the availability of cost-effective sensing solutions that allow both the creation of knowledge bases and the automatic process of them, be it with algorithmic approaches or artificial intelligence solutions. The foundation of these solutions is given by the Internet of Things (IoT), and the substanding Wireless Sensor Networks (WSN) technology stack. Of course, design approaches are needed that enable defining efficient and effective sensing infrastructures, including energy related aspects. In this paper we present a Domain Specific Language for the design of energy aware WSN IoT solutions, that allows domain experts to define sensor network models that may be then analyzed by simulation-based or analytic techniques to evaluate the effect of task allocation and offloading and energy harvesting and utilization in the network. The language has been designed to leverage the SIMTHESys modeling framework and its multiformalism modeling evaluation features.

Currently, research in Artificial Intelligence, both in Machine Learning and Deep Learning, paves the way for promising innovations in several areas. In healthcare, especially, where large amounts of quantitative and qualitative data are transferred to support studies and early diagnosis and monitoring of any diseases, potential security and privacy issues cannot be underestimated. Federated learning is an approach where privacy issues related to sensitive data management can be significantly reduced, due to the possibility to train algorithms without exchanging data. The main idea behind this approach is that learning models can be trained in a distributed way, where multiple devices or servers with decentralized data samples can provide their contributions without having to exchange their local data. Recent studies provided evidence that prototypes trained by adopting Federated Learning strategies are able to achieve reliable performance, thus by generating robust models without sharing data and, consequently, limiting the impact on security and privacy. This work propose a literature overview of Federated Learning approaches and systems, focusing on its application for healthcare. The main challenges, implications, issues and potentials of this approach in the healthcare are outlined.

Authorship Attribution (AA) is currently applied in several applications, among which fraud detection and anti-plagiarism checks: this task can leverage stylometry and Natural Language Processing techniques. In this work, we explored some strategies to enhance the performance of an AA task for the automatic detection of false and misleading information (e.g., fake news). We set up a text classification model for AA based on stylometry exploiting recurrent deep neural networks and implemented two learning tasks trained on the same collection of fake and real news, comparing their performances: one is based on Federated Learning architecture, the other on a centralized architecture. The goal was to discriminate potential fake information from true ones when the fake news comes from heterogeneous sources, with different styles. Preliminary experiments show that a distributed approach significantly improves recall with respect to the centralized model. As expected, precision was lower in the distributed model. This aspect, coupled with the statistical heterogeneity of data, represents some open issues that will be further investigated in future work.

Federated Learning is a valuable instrument for building AI-based systems that preserve the privacy and security of sensitive data, based on the main concept of shifting no more the data to the edges but moving computations to data, avoiding the collection, sharing, and use of such data by third parties. More robust federated learning systems should be able of preventing malicious inference over both data exchanged during training and the final trained model while ensuring the resulting model also has acceptable predictive accuracy. This study proposes a preliminary analysis to investigate and evaluate the effectiveness and efficiency of a federated approach to ensure valid classification accuracy and data security. A real case study from the ANDROIDS project, concerning the application of machine learning-based systems for supporting mental-health disorders detection, was considered. Large amounts of sensitive patient information are collected, which must be obfuscated or anonymized to provide a preliminary level of protection. Unfortunately, the real bottleneck lies in the difficulty of extracting all sensitive data for anonymization, due to a lot of data to handle as well as the considerable effort required. We propose a Natural Language Processing approach for sensitive knowledge detection and classification, performed by adopting a federated approach. Accuracy decay and latency introduced by applying a decentralized learning approach compared to the same task and data performed in a centralized way were evaluated. Preliminary results proved that effectiveness can be reached by a correct tuning of the federated algorithm and by choosing the right number of participants to the federation.

Natural Language Processing (NLP) is being recently explored also to its application in supporting malicious activities and objects detection. Furthermore, NLP and Deep Learning have become targets of malicious attacks too. Very recent researches evidenced that adversarial attacks are able to affect also NLP tasks, in addition to the more popular adversarial attacks on deep learning systems for image processing tasks. More precisely, while small perturbations applied to the data set adopted for training typical NLP tasks (e.g., Part-of-Speech Tagging, Named Entity Recognition, etc..) could be easily recognized, models poisoning, performed by the means of altered data models, typically provided in the transfer learning phase to a deep neural networks (e.g., poisoning attacks by word embeddings), are harder to be detected. In this work, we preliminary explore the effectiveness of a poisoned word embeddings attack aimed at a deep neural network trained to accomplish a Named Entity Recognition (NER) task. By adopting the NER case study, we aimed to analyze the severity of such a kind of attack to accuracy in recognizing the right classes for the given entities. Finally, this study represents a preliminary step to assess the impact and the vulnerabilities of some NLP systems we adopt in our research activities, and further investigating some potential mitigation strategies, in order to make these systems more resilient to data and models poisoning attacks.

Recent years have seen the widespread adoption of Artificial Intelligence techniques in several domains, including healthcare, justice, assisted driving and Natural Language Processing (NLP) based applications (e.g., the Fake News detection). Those mentioned are just a few examples of some domains that are particularly critical and sensitive to the reliability of the adopted machine learning systems. Therefore, several Artificial Intelligence approaches were adopted as support to realize easy and reliable solutions aimed at improving the early diagnosis, personalized treatment, remote patient monitoring and better decision-making with a consequent reduction of healthcare costs. Recent studies have shown that these techniques are venerable to attacks by adversaries at phases of artificial intelligence. Poisoned data set are the most common attack to the reliability of Artificial Intelligence approaches. Noise, for example, can have a significant impact on the overall performance of a machine learning model. This study discusses the strength of impact of noise on classification algorithms. In detail, the reliability of several machine learning techniques to distinguish correctly pathological and healthy voices by analysing poisoning data was evaluated. Voice samples selected by available database, widely used in research sector, the Saarbruecken Voice Database, were processed and analysed to evaluate the resilience and classification accuracy of these techniques. All analyses are evaluated in terms of accuracy, specificity, sensitivity, F1-score and ROC area.

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.

In last years smartphone and tablet devices have been handling an increasing variety of sensitive resources. As a matter of fact, these devices store a plethora of information related to our every-day life, from the contact list, the received email, and also our position during the day (using not only the GPS chipset that can be disabled but only the Wi-Fi/mobile connection it is possible to discover the device geolocalization). This is the reason why mobile attackers are producing a large number of malicious applications targeting Android (that is the most diffused mobile operating system), often by modifying existing applications, which results in malware being organized in families, where each application belonging to the same family exhibit the same malicious behaviour. These behaviours are typically information gathering related, for instance a very widespread malicious behaviour in mobile is represented by sending personal information (as examples: the contact list, the received and send SMSs, the browser history) to a remote server managed by the attackers. In this paper, we investigate whether deep learning algorithms are able to discriminate between malicious and legitimate Android samples. To this end, we designed a method based on convolutional neural network applied to syscalls occurrences through dynamic analysis. We experimentally evaluated the built deep learning classifiers on a recent dataset composed of 7100 real-world applications, more than 3000 of which are widespread malware belonging to several different families in order to test the effectiveness of the proposed method, obtaining encouraging results.

The relationships existing between a business process and the supporting software system is a critical concern for the organizations, as it directly affects their performance. The research described in this paper is concerned with the use of information retrieval techniques to software maintenance and, in particular, to the problem of recovering traceability links between the business process models and the components of the supporting software system.