A hybrid artificial intelligence (Hybrid AI) which represents a convergence of a classical (symbolic) AI with recent machine learning approaches has become a very quickly developing research axis. The combination of rule-based reasoning and statistical learning is required whenever the domain knowledge has to be incorporated in the decision system. In this work we present a system on the basis of Deep Neural Networks (DNNs) as object detectors, such as You Only Look Once version 8 (YOLOv8), transformers and logical rules which link objects and their context in the problem of rail line defect detection. Fusion of information is performed at the intermediate level - in the feature space, mixing sets of elements of this space delimited due to the object and context element detectors. Combination of objects and context elements is performed accordingly to the domain-defined rules, and fusion is ensured by a vision transformer. Experiments have been conducted on the domainrecorded dataset of rail defects. The proposed hybrid system outperforms base-line objects detection up to 0.28 of accuracy increase.

Explainable Artificial Intelligence (XAI) has become increasingly vital for improving the transparency and reliability of neural network decisions. Transformer architectures have emerged as the state-of-the-art for various tasks across single modalities such as video, language, or signals, as well as for multimodal approaches. Although XAI methods for transformers are available, their potential impact during model training remains underexplored. Thus, we propose Explanation-guided Training (ET), leveraging an XAI method to identify salient input regions and guide the model to focus solely on these salient regions during training. We develop ET in a typical multimodal analysis framework using a multimodal transformer that operates on videos and signals. ET enhances the input by masking the non-salient regions for videos and enhances the signals with weights based on explanation scores for the sensor modality. Comparative evaluation with baseline vanilla training and the state-of-the-art XAI-based IFI method [1] shows that ET consistently outperforms them. We benchmark our method on the publicly available UCF50 video dataset to demonstrate that ET is better than vanilla training and IFI. A risk detection corpus comprising egocentric videos and wearable sensor data is used for multimodal evaluation. Our code is available at https://gitub.u-bordeaux.fr/mayyar/explain-to-train

The current paradigm of training deep neural networks relies on large, annotated and representative datasets. They assume a static world where the target domain does not change. However, in the real-world, data changes over time and is often available on the fly. Naive retraining on new data causes catastrophic forgetting and the network is unable to generalize on old data. Streaming learning is a type of incremental learning where networks learn sequentially and as soon as a sample is available from the data stream. Instead of training on every new sample, we propose an uncertainty based selection criteria to improve our previously proposed fast streaming learning method Move-to-Data (MTD), called Entropy-based MTD (EMTD). Besides, streaming learning methods have so far mostly used Convolutional Neural Networks (CNNs) but in recent times Vision Transformers (ViTs) have shown much better performances for many vision tasks. Therefore, we use ViT based Video Transformer to analyse MTD, EMTD and their gradient descent based "retargeting" steps. We have compared the performances of EMTD with MTD (w/wo retargeting) and a popular streaming learning method ExStream for the transformer. EMTD is able to outperform baseline MTD, and EMTD with retargeting achieves close results as ExStream and is ∼ 1.2 times faster.

In this paper, we present the main steps of the methodology we followed to design and prototype a module dedicated to artificial intelligence (AI) for Defence for students at the level 6 in the European Qualifications Framework (EQF). Based on a top-down approach for the design taking into account the market needs, this module is based on seasonal schools, project-based learning and invited speakers coming from the local ecosystem sharing their experiences with AI. This combination aims at attracting students to engineering in the field of Defence. Finally, it should be noted that a part of this module can be done in a hybrid way.

In Deep Neural Network training, the availability of a large amount of representative training data is the sine qua non-condition for a good generalization capacity of the model. In many real-world applications, data is not available at a glance, but coming on the fly. If a pre-trained model is fine-tuned on the new data, then catastrophic forgetting happens mostly. Incremental learning mechanisms propose ways to overcome catastrophic forgetting. Streaming learning is a type of incremental learning where models learn from new data instances as soon as they become available in a single training pass. In this work, we conduct an experimental study, on a large dataset, of an incremental/streaming learning method Move-to-Data we previously proposed, and propose an updated approach by ”re-targeting” with gradient descent which is faster than the popular streaming learning method ExStream. The method achieves better performances and computational efficiency compared to ExStream. Move-to-Data with gradient is on average 3.5 times faster than ExStream and has a similar accuracy, with 0.5% improvement compared to ExStream.

Recently, visual transformers have shown promising results in tasks such as image classification, segmentation, object detection, etc. The explanation of their decision remains a challenge. This paper focuses on exploiting self-attention for an explanation. We propose a generalized interpretation of the transformers i.e model agnostic but class-specific explanations. The main principle is in the use and weighting self-attention maps of a visual transformer. To evaluate it, we use the popular hypothesis that an explanation is good if it correlates with human perception of a visual scene. Thus, the method has been evaluated against the Gaze Fixation Density Maps obtained in a psycho-visual experiment on a public database. It has been compared with other popular explainers such as Grad-Cam, LRP, Rollout, and Adaptive Relevance methods. The proposed method outperforms the best baseline by 2% in a standard Pearson Correlation Coefficient (PCC) metric.

The paper proposes a video transformer architecture for detection of risk events on frail adults with ego video monitoring data. First we introduce an extended taxonomy for risk events, and then we propose a transformer based video recognition model for detection of these risk events. The proposed transformer architecture consists of separable attention for spatial and temporal data. We also introduce a pooling operation on the temporal video data by learning of their importance. The experiments have been conducted on visual data of in-the-wild recorded BIRDS dataset and on Kinetics-400 for benchmarking. The use of the pooling operation in transformers gives an increment of 3% on BIRDS dataset.

As more and more data is generated and stored, and as longer data streams become available, concept drift detection is becoming crucial for most real world applications. We introduce Partially Supervised Drift Detection, PSDD, a drift detection method based on Decision Trees that does not suppose any knowledge of true class labels during inference. Our approach works in any number of dimensions and is able to distinguish real from virtual drift. We successfully evaluated our method with well established datasets in the drift detection field.

Recent years have observed an exponential growth in the popularity of audio-based authentication systems. The benefit of a voice-based authentication system is that the person need not be physically present. Voice biometric system provides effective authentication in various domains like remote access control, authentication in mobile applications, customer care centers for call attests. Most of the existing authentication systems that recognize speakers formulate deep learning models for better classification. At the same time, research studies show that deep learning models are highly vulnerable to adversarial inputs. A breach in security on authentication systems are not generally acceptable. This paper exposes the vulnerabilities of audio-based authentication systems. Here, we propose a novel downsampling attack to the speaker recognition system. This attack can effectively trick the speaker recognition framework by causing inaccurate predictions. The proposed threat model achieved remarkable attack effectiveness of 75%. This system employs a custom human voice dataset recorded in real-time conditions to achieve real-time effectiveness during classification. We compare the attack accuracy of the proposed attack against the adversarial audios generated using the CleverHans toolbox. The proposed attack being a black box attack, is transferable to other deep learning systems also.

Multimedia today is also in multimodality. Working with heterogeneous signals we use multimedia techniques of data fusion and mining. Classification from real world datasets are often challenging. The paper is devoted to the detection of personal risk situations of fragile people from multi-modal sensing real world lifelog data named BIRDS. Using a real-world data is challenging as the risk situations are rare and last just a few seconds compared to the global volume of the dataset. In this paper we propose a GRU architecture with global attention block to recognise semantic risk situations from a limited taxonomy. Attention is also focused on data organisation and pre-processing with imputation and normalisation. The proposed method is applied to a real-world collected multimodal dataset and to the OpenSource dataset UCI-HAR for the sake of comparison with the state-of-the-art.

Classification of Alzheimer’s disease from 3D structural Magnetic Resonance Imaging (sMRI) with deep neural networks has shown promising results in recent years. The decision interpretation of these networks is essential to aid medical experts to understand and rely on the results provided by such models. In this paper, we propose an adaptation of a recently developed feature-based explanation method and apply it to a 3D CNN architecture for the binary classification of Alzheimer’s disease and Normal Control from the hippocampal ROIs of brain sMRIs. We also compare our method to the state-of-the-art LRP method.

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%.

Data theft is a significant security threat for mobile app users. The growing importance of digitization motivates the diversity of available applications. In this paper, we propose a novel and lightweight method for classifying Android apps into low, medium, and high-risk categories. Our approach relies largely on the other permissions (also termed as hidden permissions) of the Android applications. We have proposed a linear regression-based technique to classify the apps into different risk categories. We will show how other permissions can be used as a strong indicator for defining risk categories. We have used K-means clustering to validate and explain the decision of our method. In an evaluation with 500 applications and 101 other permissions, our proposed approach decides the risk factor of an app, and the explanation is provided for each detection reveal relevant properties of the detected risk.

As Internet of Things (IoT) is widely spread and is becoming heterogeneous, a growing number of connected devices are the focus of security threats. Hence, a standardized security strategy seems required. OneM2M [1] is a global standard initiative designed to satisfy the need for a common horizontal platform for the multi-industry M2M/IoT applications. In this paper, we propose an Intrusion Detection and Prevention System (IDPS), for the Service Layer introduced by the oneM2M standard. To our knowledge, it is the first generic IDPS for the oneM2M Service Layer based on Edge Machine Leaning (ML). We will detail, in this work, the strategy of the oneM2M-IDPS. Moreover, we investigate the performance of ML algorithms on the oneM2M generated dataset to choose the best ones for our IDPS. Since we are in the context of tiny devices (IoT), we pay attention in our experiments to the features dimension reduction in ML and thus, to the size of trained models.

Cut-in into vehicle platoons is a dangerous driving maneuver that affects the safety and efficiency of platooning vehicles. An accurate prediction of such maneuver enables the platooning system to take safety measures that ensure the platoon safety and integrity. The contribution of this paper consists of an evaluation of a set of supervised machine learning algorithms for cut-in maneuver recognition, for eventual use in platooning systems. The models were trained and tested on a large-scale publicly available driving dataset, from which cut-in events were extracted. The results show that tree-based classifiers such as Gradient Booting Machine can recognize the cut-in maneuvers with an Fl-score of 98%. An experiment to investigate the model performance with advanced prediction times shows that up to 80.5% of the cut-ins were correctly predicted 1 second before the lane crossing time.

In the present work, we investigate the possibility to expand existing Deep Learning solvers in order to improve the quality and stability of the training. We propose different new solvers, all of them based on the filtering of the neural network parameters, and experimentally prove that properly tuning their respective hyper-parameters leads to a clear improvement of the training and validation results, in quality and in stability.

Wireless Body Area Networks are nowadays attracting both academic and industrial worlds. Combining collected data related to patient context with original health measurement can enhance the general health state monitoring and help to better understand the patient disease evolution. Daily activity is one of the important features that may influence the patient health state. Thus, recognizing the user activity can be a useful way for improving quality of health services. Relying on supervised learning, we study the feasibility of extracting and classifying the human activities from channel gain measures, which is an important feature that characterizes the WBAN channel links.

Sensitive (private) information can escape from one app to another using one of the multiple communication methods provided by Android for inter-app communication. This leakage can be malicious. In such a scenario, individual benign app, in collusion with other conspiring apps, if present, can leak the private information. In this work in progress, we present, a new model-checking based approach for inter-app collusion detection. The proposed technique takes into account simultaneous analysis of multiple apps. We are able to identify any set of conspiring apps involved in the collusion. To evaluate the efficacy of our tool, we developed Android apps that exhibit collusion through inter-app communication. Eight demonstrative sets of apps have been contributed to widely used test dataset named DroidBench. Our experiments show that proposed technique can accurately detect the presence/absence of collusion among apps. To the best of our knowledge, our proposal has improved detection capability than other techniques.

Android applications need to access and share user's sensitive data. To maintain user's privacy and related data security, it is essential to protect this data. Android security framework enforces permission protected model but it has been shown that applications can bypass this security model. Attacks based on such unauthorized privileges are known as Inter-Component Communication (ICC) Collusion Attacks. In this paper, we propose, a novel automaton framework that allows effective detection of intent based collusion. Our detection framework operates at the component-level. To evaluate our proposal, we developed 14 applications and took 4 applications from Google Play Store. We took all possible combinations from the set of 21 applications. We tested our approach on 210 pairs of applications derived from the set of 21 applications. Time and space complexity of our proposed approach isO(n) where n is the number of components in all the applications under analysis. The experimental results demonstrate that our technique is scalable to application sizing and more efficient as compared to other state of the art approaches.

We consider networks of processes which interact with beeps. In the basic model defined by Cornejo and Kuhn, which we refer to as the BL variant, processes can choose in each round either to beep or to listen. Those who beep are unable to detect simultaneous beeps. Those who listen can only distinguish between silence and the presence of at least one beep. Stronger variants exist where the nodes can also detect collision while they are beeping (B_{cd}L) or listening (BL_{cd}), or both (B_{cd}L_{cd}). Beeping models are weak in essence and even simple tasks are difficult or unfeasible with them. This paper starts with a discussion on generic building blocks (design patterns) which seem to occur frequently in the design of beeping algorithms. They include multi-slot phases: the fact of dividing the main loop into a number of specialised slots; exclusive beeps: having a single node beep at a time in a neighbourhood (within one or two hops); adaptive probability: increasing or decreasing the probability of beeping to produce more exclusive beeps; internal (resp. peripheral) collision detection: for detecting collision while beeping (resp. listening); and emulation of collision detection: for enabling this feature when it is not available as a primitive. We then provide algorithms for a number of basic problems, including colouring, 2-hop colouring, degree computation, 2-hop MIS, and collision detection (in BL). Using the patterns, we formulate these algorithms in a rather concise and elegant way. Their analyses (in the full version) are more technical, e.g. one of them relies on a Martingale technique with non-independent variables; another improves that of the MIS algorithm (P. Jeavons et al.) by getting rid of a gigantic constant (the asymptotic order was already optimal). Finally, we study the relative power of several variants of beeping models. In particular, we explain how every Las Vegas algorithm with collision detection can be converted, through emulation, into a Monte Carlo algorithm without, at the cost of a logarithmic slowdown. We prove that this slowdown is optimal up to a constant factor by giving a matching lower bound.

Android being the most popular open source mobile operating system, attracts a plethora of app developers. Millions of applications are developed for Android platform with a great extent of behavioral diversities and are available on Play Store as well as on many third party app stores. Due to its open nature, in the past Android Platform has been targeted by many malware writers. The conventional way of signature-based detection methods for detecting malware on a device are no longer promising due to an exponential increase in the number of variants of the same application with different signatures. Moreover, they lack in dynamic analysis too. In this paper, we propose DRACO, which employs a two-phase detection technique that blends the synergy of both static and dynamic analysis. It has two modules, client module that is in the form an Android app and gets installed on mobile devices and a server module that runs on a server. DRACO also explains user about the features contributing to the maliciousness of analyzed app and generates scoring for that maliciousness. It does not require any root or super-user privileges. In an evaluation of 18,000 benign applications and 10,000 malware samples, DRACO performs better than several related existing approaches and detects 98.4% of the malware with few false alerts. On ten popular smartphones, the method requires an average of 6 seconds for on device analysis and 90 seconds on server analysis.

In Wireless Sensor Networks, flooding in one of the basic communication primitives. It is used to propagate informations from one node to the entire network. Every node, receiving a piece of information, has to flood it to all its neighborhood. As informations to spread are important, such as fire or intrusion alerts, communications should be secured. In this paper, we present a greedy flooding algorithm for Wireless Sensor Networks secured by a Random Key Pre-distribution model that make them redoubtable. We present two theoretical upper bounds of the time complexity of this algorithm that depend on the network structure. We then use the ViSiDiA platform to implement and simulate our algorithm to validate these theoretical results.

The problem of providing source authentication in Wireless Sensor Networks (WSNs) has been a roadblock to their large scale deployment, and is still in its infancy. In this paper, we present novel symmetric-key-based authentication schemes which exhibit low computation and communication authentication overhead. Our schemes are built upon the integration of a reputation mechanism, a Bloom filter, and a key binary tree for the distribution and updating of the authentication keys. Analytical evaluation of the proposed authentication schemes shows that the estimated average number of concatenated message authentication code in a packet from time 0 till time t is 4pt, with p is the probability that a key is corrupted. Our schemes are lightweight and efficient with respect to computation, communication and energy overhead.

SUMMARY Wireless sensor networks are a new generation of networks that need specific models and algorithms. We are interested specifically in mobile wireless sensor networks that are considered as anonymous asynchronous distributed mobile systems. As broadcast is one of the most important applications for such networks, and as it depends on the communication model, we tried to find the most suitable one to make a distributed broadcast algorithm. We adopted the population protocols, the Angluin's model of pairwise interactions of anonymous finite‐state agents, to broadcast an information. We tried to modify this model to avoid the information duplication and then calculated the complexity of the algorithm. Then, we extended the model with the rendezvous one that made the stabilization of the algorithm faster. The implementation, the simulation, and the validation of these algorithms and results have been done with Visidia. Copyright © 2012 John Wiley & Sons, Ltd.

This paper presents efficient deterministic and randomized distributed algorithms for decomposing a graph with n nodes into a disjoint set of connected clusters with small radius and few intercluster edges. Our algorithms can be easily implemented in the distributed CONGEST model of computation i.e., limited message size, improving the time complexity of previous algorithms (Moran and Snir, 2000; Awerbuch, 1985; Peleg, 2000) from linear to sublinear. One important application of our algorithms is efficient construction of sparse graph spanners. In fact, given a parameter k, we show that there exists a sublinear deterministic distributed algorithm that constructs a graph spanner of stretch 2k - 1 with at most O(n1+1k/) edges in the CONGEST model