Network intrusion detection systems play a crucial role in the security strategy employed by organisations to detect and prevent cyberattacks. Such systems usually combine pattern detection signatures with anomaly detection techniques powered by machine learning methods. However, the commonly proposed machine learning methods present drawbacks such as over-reliance on labeled data and limited generalization capabilities. To address these issues, embedding-based methods have been introduced to learn representations from network data, such as DNS traffic, mainly due to its large availability, that generalise effectively to many downstream tasks. However, current approaches do not properly consider contextual information among DNS queries. In this paper, we tackle this issue by proposing DNS-GT, a novel Transformer-based model that learns embeddings for domain names from sequences of DNS queries. The model is first pre-trained in a self-supervised fashion in order to learn the general behavior of DNS activity. Then, it can be finetuned on specific downstream tasks, exploiting interactions with other relevant queries in a given sequence. Our experiments with real-world DNS data showcase the ability of our method to learn effective domain name representations. A quantitative evaluation on domain name classification and botnet detection tasks shows that our approach achieves better results compared to relevant baselines, creating opportunities for further exploration of large-scale language models for intrusion detection systems. Our code is available at: https://github.com/m-altieri/DNS-GT.

Network science has been a rapidly evolving field to study systems made of interactions between entities. Studying the structure of such networks reveals indeed the underlying mechanisms of these systems, and has been proven successful in many domains, such as sociology, biology, or geography. Recently, connections between network science and signal processing have emerged, making the use of a wide variety of tools possible to study networks. In this chapter, a focus is made on a methodology introduced to transform a graph into a collection of signals, using a multidimensional scaling technique: by projecting a distance matrix representing relations between vertices of the graph as points in a Euclidean space, it is possible to interpret coordinates of vertices in this space as signals, and take advantage of this dual representation to develop new tools for the study of networks. Deeper considerations of this methodology are proposed, by strengthening the connections between the obtained signals and the common graph structures. A robust inverse transformation method is next described, taking into account possible changes in the signals. Establishing a robust duality between graphs and signals opens up new perspectives, as classical signal processing tools, such as spectral analysis or filtering, are made available for the study of the structure of networks.

This study explores the complexities of integrating Artificial Intelligence (AI) into Autonomous Vehicles (AVs), examining the challenges introduced by AI components and the impact on testing procedures, focusing on some of the essential requirements for trustworthy AI. Topics addressed include the role of AI at various operational layers of AVs, the implications of the EU's AI Act on AVs, and the need for new testing methodologies for Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS). The study also provides a detailed analysis on the importance of cybersecurity audits, the need for explainability in AI decision-making processes and protocols for assessing the robustness and ethical behaviour of predictive systems in AVs. The paper identifies significant challenges and suggests future directions for research and development of AI in AV technology, highlighting the need for multidisciplinary expertise.

This article examines the interplay between artificial intelligence (AI) and cybersecurity in light of future regulatory requirements on the security of AI systems, specifically focusing on the robustness of high-risk AI systems against cyberattacks in the context of the European Union’s AI Act. The paper identifies and analyses three challenges to achieve compliance of AI systems with the cybersecurity requirement: accounting for the diversity and the complexity of AI technologies, assessing AI-specific risks, and developing secure-by-design AI systems. The contribution of the article consists in providing an overview of AI cybersecurity practices and identifying gaps in current approaches to security conformity assessment for AI systems. Our analysis highlights the unique vulnerabilities present in AI systems and the absence of established cybersecurity practices tailored to these systems, and emphasises the need for continuous alignment between legal requirements and technological capabilities, acknowledging the necessity for further research and development to address the challenges. It concludes that comprehensive cybersecurity practices must evolve to accommodate the unique aspects of AI, with a collaborative effort from various sectors to ensure effective implementation and standardisation.

The proposed EU regulation for Artificial Intelligence (AI), the AI Act, has sparked some debate about the role of explainable AI (XAI) in high-risk AI systems. Some argue that black-box AI models will have to be replaced with transparent ones, others argue that using XAI techniques might help in achieving compliance. This work aims to bring some clarity as regards XAI in the context of the AI Act and focuses in particular on the AI Act requirements for transparency and human oversight. After outlining key points of the debate and describing the current limitations of XAI techniques, this paper carries out an interdisciplinary analysis of how the AI Act addresses the issue of opaque AI systems. In particular, we argue that neither does the AI Act mandate a requirement for XAI, which is the subject of intense scientific research and is not without technical limitations, nor does it ban the use of black-box AI systems. Instead, the AI Act aims to achieve its stated policy objectives with the focus on transparency (including documentation) and human oversight. Finally, in order to concretely illustrate our findings and conclusions, a use case on AI-based proctoring is presented.

Adversarial machine learning (AML), by designing attacks that intentionally break or misuse state-of-the-art machine learning models, has become the most prominent scientific field to explore the security aspects of Artificial Intelligence. A whole range of vulnerabilities, previously irrelevant in traditional ICT, have effectively emerged in these studies. In the light of upcoming legislations mandating security requirements for AI products and services, there is a need to understand how AML techniques connect with the broader field of cybersecurity, and how to articulate more tightly threat models with realistic cybersecurity procedures. This article aims to contribute to closing the gap between AML and cybersecurity by proposing an approach to study the feasibility of an attack in a cybersecurity risk assessment framework, illustrated with a specific use case of an evasion attack designed to fool traffic sign recognition systems in the physical world. The importance of considering the feasibility of carrying out such attacks under real conditions is emphasized through the analysis of two factors: the reproducibility of the attack according to a published description or existing code, and the applicability of the attack by a malicious actor operating in a real-world environment.

New emerging technologies powered by Artificial Intelligence (AI) have the potential to disruptively transform our societies for the better. In particular, data-driven learning approaches (i.e., Machine Learning (ML)) have been a true revolution in the advancement of multiple technologies in various application domains. But at the same time there is growing concern about certain intrinsic characteristics of these methodologies that carry potential risks to both safety and fundamental rights. Although there are mechanisms in the adoption process to minimize these risks (e.g., safety regulations), these do not exclude the possibility of harm occurring, and if this happens, victims should be able to seek compensation. Liability regimes will therefore play a key role in ensuring basic protection for victims using or interacting with these systems. However, the same characteristics that make AI systems inherently risky, such as lack of causality, opacity, unpredictability or their self and continuous learning capabilities, may lead to considerable difficulties when it comes to proving causation. This paper presents three case studies, as well as the methodology to reach them, that illustrate these difficulties. Specifically, we address the cases of cleaning robots, delivery drones and robots in education. The outcome of the proposed analysis suggests the need to revise liability regimes to alleviate the burden of proof on victims in cases involving AI technologies.

Can satisfactory explanations for complex machine learning models be achieved in high-risk automated decision-making? How can such explanations be integrated into a data protection framework safeguarding a right to explanation? This article explores from an interdisciplinary point of view the connection between existing legal requirements for the explainability of AI systems set out in the General Data Protection Regulation (GDPR) and the current state of the art in the field of explainable AI. It studies the challenges of providing human legible explanations for current and future AI-based decision-making systems in practice, based on two scenarios of automated decision-making in credit scoring risks and medical diagnosis of COVID-19. These scenarios exemplify the trend towards increasingly complex machine learning algorithms in automated decision-making, both in terms of data and models. Current machine learning techniques, in particular those based on deep learning, are unable to make clear causal links between input data and final decisions. This represents a limitation for providing exact, human-legible reasons behind specific decisions, and presents a serious challenge to the provision of satisfactory, fair and transparent explanations. Therefore, the conclusion is that the quality of explanations might not be considered as an adequate safeguard for automated decision-making processes under the GDPR. Accordingly, additional tools should be considered to complement explanations. These could include algorithmic impact assessments, other forms of algorithmic justifications based on broader AI principles, and new technical developments in trustworthy AI. This suggests that eventually all of these approaches would need to be considered as a whole.

Can we achieve an adequate level of explanation for complex machine learning models in high-risk AI applications when applying the EU data protection framework? In this article, we address this question, analysing from a multidisciplinary point of view the connection between existing legal requirements for the explainability of AI systems and the current state of the art in the field of explainable AI. We present a case study of a real-life scenario designed to illustrate the application of an AI-based automated decision making process for the medical diagnosis of COVID-19 patients. The scenario exemplifies the trend in the usage of increasingly complex machine-learning algorithms with growing dimensionality of data and model parameters. Based on this setting, we analyse the challenges of providing human legible explanations in practice and we discuss their legal implications following the General Data Protection Regulation (GDPR). Although it might appear that there is just one single form of explanation in the GDPR, we conclude that the context in which the decision-making system operates requires that several forms of explanation are considered. Thus, we propose to design explanations in multiple forms, depending on: the moment of the disclosure of the explanation (either ex ante or ex post); the audience of the explanation (explanation for an expert or a data controller and explanation for the final data subject); the layer of granularity (such as general, group-based or individual explanations); the level of the risks of the automated decision regarding fundamental rights and freedoms. Consequently, explanations should embrace this multifaceted environment. Furthermore, we highlight how the current inability of complex, deep learning based machine learning models to make clear causal links between input data and final decisions represents a limitation for providing exact, human-legible reasons behind specific decisions. This makes the provision of satisfactorily, fair and transparent explanations a serious challenge. Therefore, there are cases where the quality of possible explanations might not be assessed as an adequate safeguard for automated decision-making processes under Article 22(3) GDPR. Accordingly, we suggest that further research should focus on alternative tools in the GDPR (such as algorithmic impact assessments from Article 35 GDPR or algorithmic lawfulness justifications) that might be considered to complement the explanations of automated decision-making.

................................................................................................................................................................................................................................................................3 Acknowledgements..................................................................................................................................................................................................................................... 4 Executive summary .................................................................................................................................................................................................................................... 5

When using optimization methods with matrix variables in signal processing and machine learning, it is customary to assume some low-rank prior on the targeted solution. Nonnegative matrix factorization of spectrograms is a case in point in audio signal processing. However, this low-rank prior is not straightforwardly related to complex matrices obtained from a short-time Fourier - or discrete Gabor - transform (STFT), which is generally defined from and studied based on a modulation operator and a translation operator applied to a so-called window. This paper is a first study of the low-rankness property of time-frequency matrices. We characterize the set of signals with a rank-r (complex) STFT matrix in the case of a unit hop size and frequency step with few assumptions on the transform parameters. We discuss the scope of this result and its implications on low-rank approximations of STFT matrices.

Musical noise is a recurrent issue that appears in spectral techniques for denoising or blind source separation. Due to localised errors of estimation, isolated peaks may appear in the processed spectrograms, resulting in annoying tonal sounds after synthesis known as “musical noise”. In this paper, we propose a method to assess the amount of musical noise in an audio signal, by characterising the impact of these artificial isolated peaks on the processed sound. It turns out that because of the constraints between STFT coefficients, the isolated peaks are described as time-frequency “spots” in the spectrogram of the processed audio signal. The quantification of these “spots”, achieved through the adaptation of a method for localisation of significant STFT regions, allows for an evaluation of the amount of musical noise. We believe that this will pave the way to an objective measure and a better understanding of this phenomenon.

A new framework to track the structure of temporal networks with a signal processing approach is introduced. The method is based on the duality between static networks and signals, obtained using a multidimensional scaling technique, that makes possible the study of the network structure from frequency patterns of the corresponding signals. In this paper, we propose an approach to identify structures in temporal networks by extracting the most significant frequency patterns and their activation coefficients over time, using non-negative matrix factorization of the temporal spectra. The framework, inspired by audio decomposition, allows transforming back these frequency patterns into networks, to highlight the evolution of the underlying structure of the network over time. The effectiveness of the method is first evidenced on a synthetic example, prior being used to study a temporal network of face-to-face contacts. The extracted subnetworks highlight significant structures decomposed on time intervals that validates the relevance of the approach on real-world data.

In recent years, bike sharing systems have become very popular in many major 1 cities. Thanks to the data they generate, their activity can be tracked down, giving 2 an overall view of how human activities are spread over time and space. We propose 3 in the present article a novel method to extract mobility patterns that occur in such 4 large-scale transportation systems. The trips made by the users are first represented 5 as flows between the different stations of the system, describing a network whose 6 structure evolves over time. A decomposition technique is then proposed using 7 non-negative matrix factorisation, to express the resulting temporal networks as 8 a mixture of sub-networks, each of them characterising the different behaviours 9 of users over time and space. This method is applied on the Lyon’s bike sharing 10 system, and it is emphasised that key spatio-temporal elements of urban activity 11 are retrieved, capturing known phenomena such as commuting. This approach 12 could be easily extended to large-scale transportation systems exhibiting a network 13 structure, paving the way to an unsupervised modelling of mobility patterns. 14

Convex nonnegative matrix factorization (CNMF) is a variant of nonnegative matrix factorization (NMF) in which the components are a convex combination of atoms of a known dictionary. In this contribution, we propose to extend CNMF to the case where the data matrix and the dictionary have missing entries. After a formulation of the problem in this context of missing data, we propose a majorization-minimization algorithm for the solving of the optimization problem incurred. Experimental results with synthetic data and audio spectro-grams highlight an improvement of the performance of reconstruction with respect to standard NMF. The performance gap is particularly significant when the task of reconstruction becomes arduous, e.g. when the ratio of missing data is high, the noise is steep, or the complexity of data is high.

L'analyse archetypale (AA), ou factorisation convexe en matrices non-negatives (CNMF), est une variante de la factorisation en matrices non-negatives (NMF), dans laquelle les composantes obtenues sont exprimees comme une combinaison convexe d'exemples appeles archetypes. Dans cette contribution, nous proposons d'etendre AA/CNMF au cas ou la matrice des donnees et la matrice des archetypes sont partiellement observees. Apres avoir reformule le probleme dans ce contexte de donnees manquantes, nous proposons un algorithme de type Majorisation-Minimisation pour l'estimation des facteurs de la decomposition puis la reconstruction des donnees manquantes. Une comparaison est realisee sur des donnees synthetiques, mettant en evidence une amelioration des performances de reconstruction de donnees manquantes par rapport a la NMF classique. L'ecart de performance se revele particulierement interessant lorsque le bruit est important ou que le nombre de donnees manquantes est grand.