Archive ouverte HAL-LAAS
Not a member yet
12189 research outputs found
Sort by
Silicon-Silica Core-Shell Particles Synthesis for Directional Light Scattering
No full textInternational audienc
IA : comment les grands modèles de langage peuvent devenir des super méchants… entre de mauvaises mains
No full text[VULGARISATION SCIENTIFIQUE]Avec l’arrivée des grands modèles de langage (LLM), les attaques informatiques se multiplient. Il est essentiel de se préparer à ces LLM entraînés pour être malveillants, car ils permettent d’automatiser le cybercrime. En mai, un LLM a découvert une faille de sécurité dans un protocole très utilisé… pour lequel on pensait que les failles les plus graves avaient déjà été décelées et réparées.Pour rendre un LLM malveillant, les pirates détournent les techniques d'apprentissage à la base de ces outils d'IA et contournent les garde-fous mis en place par les développeurs.Avec l’arrivée des grands modèles de langage (LLM), les attaques informatiques se multiplient. Il est essentiel de se préparer à ces LLM entraînés pour être malveillants, car ils permettent d’automatiser le cybercrime. En mai, un LLM a découvert une faille de sécurité dans un protocole très utilisé… pour lequel on pensait que les failles les plus graves avaient déjà été décelées et réparées.Pour rendre un LLM malveillant, les pirates détournent les techniques d'apprentissage à la base de ces outils d'IA et contournent les garde-fous mis en place par les développeurs
Impact of mechanical compressive stress on pancreatic cancer progression
No full textInternational audienceContext: Mechanical compressive stress arises during pancreatic cancer progression (PDAC). In vitro, compression forces decrease PDAC cell proliferation, increase invasiveness, and induce resistance to chemotherapies. In vivo, the importance of compression is unknown. In PDAC, increased compressive stress happens simultaneously with the second wave of genetic alterations (p53 mutations/truncations) after KRAS oncogenic mutations; it is also linked with overexpression/activation of the PI3-Kinases (PI3K) pathway. We think that compression favors selective genetic backgrounds that modify the signaling environment in cells and thus cell fate. Experimental design: We generated compressive stresses to spheroids derived from PDAC cells with KRAS G12D mutation, in which p53 R172H mutation or p53 R172H;R210* truncation are induced sequentially. Further, we applied a compressive stress to KRAS G12D ±p53 R172H /p53 R172H;R210* mutated mouse allografts using a compressive device. We also used the punch method in order to evaluate the relaxation of tumors depending on their genetic background. Results: Compression decreased the spheroid growth (<30%). However, p53 R172H mutated PDAC cells developed a resistance to compression and continued to proliferate. This mutation associated with a truncation of p53 accentuated this resistance, even bringing a proliferative advantage. A transcriptomic analysis of KRAS G12D , p53 mutated and p53 truncated spheroids under compression was performed. This analysis showed a modification in adhesion properties via plasma membrane and RTK signaling activity, mechanisms regulated by PI3K pathway. In parallel, we observed, in vivo, that the growth of KRAS G12D mutated tumors decreased by 40% under compression, whereas the size of tumors with p53 R172H;R210* truncated form was similar with or without compression. Finally, KRAS G12D tumors relaxed more easily compared to the p53 R172H and p53 R172H;R210* tumors; this was due to a greater cellular and matrix homogeneity in these tumors compared to p53 R172H and p53 R172H;R210* tumors. Conclusion: Growth under pressure can influence the progression of PDAC promoting selective genetic background and activation of oncogenic signaling pathways. These observations open the way to integrate the mechanical context in the management of patients with PDAC
When the weakest model sees the threat: an explainable ensemble learning system for detecting network attacks
No full textInternational audienceThe growing importance of network security is driven by two major challenges. One is the exponential increase in network traffic, which exceeds human processing capabilities. The other is the rising frequency and sophistication of attacks, which demand advanced, intelligent analysis. To take critical actions-such as blocking suspicious IP addresses-security analysts must understand why intrusion detection systems raise alarms. This highlights the limitations of relying on machine learning models with opaque decision-making processes, often referred to as "black boxes," whose lack of interpretability poses challenges for justifying security actions. Consequently, this paper emphasizes the need for explainable machine learning solutions tailored to network security. To detect new attacks effectively, we adopt a behavioral approach that analyzes short time windows of aggregated traffic to identify abnormal patterns, using various unsupervised machine learning detectors. These detectors often yield complementary results: they may disagree on specific detections, and in some cases, a generally less effective model-the weakest-can uniquely identify an attack. This observation motivates an ensemble approach that integrates the strengths of diverse models. Our approach combines three key contributions: a stacking-based ensemble learning strategy that improves detection by incorporating minority reports, going beyond majority voting; a visual representation technique
Proportional modulation of proliferation and motility under 2D compressive stress depends on mesenchymal phenotype
No full textInternational audienceMechanical stresses, including compression, arise during cancer progression. In solid cancer, especially breast and pancreatic cancers, the rapid tumor growth and the environment remodeling explain their high intensity of compressive forces. However, the sensitivity of compressed cells to targeted therapies remains poorly known. In breast and pancreatic cancer cells, pharmacological PI3K inactivation decreased cell number and induced apoptosis. These effects were accentuated when we applied 2D compression forces in mechanically responsive cells. Compression selectively induced the overexpression of PI3K isoforms and PI3K/AKT pathway activation. Furthermore, transcriptional effects of PI3K inhibition and compression converged to control the expression of an autophagy regulator, GABARAP, whose level was inversely associated with PI3K inhibitor sensitivity under compression. Compression alone blocked autophagy flux in all tested cells, whereas inactivation of basal PI3K activity restored autophagy flux only in mechanically non-responsive compressed cells. This study provides direct evidence for the role of the PI3K/AKT pathway in compression-induced mechanotransduction. PI3K inhibition promotes apoptosis or autophagy, explaining PI3K importance to control cancer cell survival under compression
Plateforme d'analyse de signaux micro-architecturaux pour la mise en place de compteurs matériels de sécurité
No full textInternational audienceLa détection de comportements malveillants logiciels ou matériels pendant le fonctionnement d'un système informatique demande la mise en place d'observables provenant d'une ou plusieurs couches d'abstraction de ce dernier. Cette abstraction cependant tend à limiter la capacité à détecter des déviations de comportement, surtout pour des classes d'attaques qui exploitent des vulnérabilités très proches du matériel cible. A contrario, un niveau d'abstraction trop faible tend à faire croître significativement la complexité du modèle du système et donc pose un certain nombre de difficultés pour l'extraction et la sélection des observables pertinents pour une classe d'attaque donnée.Les compteurs de performance des processeurs ont notamment été utilisés comme moyen indirecte d'observer le comportement de la micro-architecture et détecter des logiciels tentant d'exploiter des vulnérabilités matérielles. Afin d'améliorer les différentes méthodes de détection, nous proposons la construction de métriques matérielles pensées dès la conception pour la sécurité en étudiant la corrélation entre les signaux provenant de la micro-architecture et les différentes classes d'attaque de la littérature ciblant à la fois des systèmes IT classiques et OT industriels.Par extension, ces travaux ont pour ambition de détecter des attaques provenant de chevaux de Troie matériels, ces derniers ayant pour effet de changer le comportement d'une micro-architecture donnée
Rethinking Cyber Safety and Cybersecurity in the Age of AI: A Position Paper
No full textInternational audienceDistinguishing between cyber safety and cybersecurity is crucial, as understanding and addressing both are essential for developing comprehensive strategies to protect systems, data, users, and their environments. In this article, we argue that the terms cyber safety and cybersecurity lack a clear distinction in research and practice. We also explore how these concepts interact and potentially influence each other, particularly in the context of generative AI
Reconsidering confidence in assurance cases: From quantification to strength-of-knowledge aggregation
No full textInternational audienceThere is no agreed method to express and evaluate confidence in assurance cases to accept or reject claims. Quantitative approaches based on probability theory or extensions thereof have been criticized based on inconsistencies and ambiguity in interpretation. Moreover, the quality of any quantitative analysis ultimately depends on the underlying knowledge basis. Thus, one cannot escape some form of qualitative assessment even when quantitative methods are used. In this paper we suggest a method for aggregating qualitative confidence assessments. We treat leaf claims as 'true', 'false' or 'uncertain' based on strengthof-knowledge criteria and use three-valued logic to propagate the knowledge strength through argument steps. The use of threevalued logic conveniently accommodates the distinct roles of sub claims and side claims, and also avoids the need to distinguish different kinds of defeaters
Gradient-based optimization of core-shell particles with discrete materials for directional scattering
No full textInternational audienceDesigning nanophotonic structures traditionally grapples with the complexities of discrete parameters, such as real materials, often resorting to costly global optimization methods. This paper introduces an approach that leverages generative deep learning to map discrete parameter sets into a continuous latent space, enabling direct gradient-based optimization. For scenarios with non-differentiable physics evaluation functions, a neural network is employed as a differentiable surrogate model. The efficacy of this methodology is demonstrated by optimizing the directional scattering properties of core-shell nanoparticles composed of a selection of realistic materials. We derive suggestions for core-shell geometries with strong forward scattering and minimized backscattering. Our findings reveal significant improvements in computational efficiency and performance when compared to global optimization techniques. Beyond nanophotonics design problems, this framework holds promise for broad applications across all types of inverse problems constrained by discrete variables
Vers une maintenance prévisionnelle interprétable par modèles d’apprentissage automatique dans les systèmes complexes à composants tournants
No full textNational audiencePredictive maintenance has become essential in the context of Industry 4.0, particularly for critical rotating components within complex industrial systems, where failures can lead to significant costs and unplanned production downtimes. However, most predictive models currently used remain opaque and are often perceived as black boxes by operators, limiting their adoption in real-world settings. This thesis focuses on the development of interpretable and explainable approaches to predictive maintenance, aiming to combine predictive performance with transparency to enhance their acceptability in industrial environments. To achieve this objective, several contributions are proposed. First, a methodology for optimizing multivariate time windows is developed to improve the structuring of non-stationary time series used in predictive maintenance. This method enables the automatic determination of optimal segmentation parameters, thereby facilitating the extraction of relevant information for training predictive models. Next, Multiclass Neural Additive Models (MNAM) are introduced as explainable machine learning models tailored to multiclass classification problems encountered in industrial contexts, such as the prediction of the Remaining Useful Life (RUL) of rotating components, framed as a classification task. MNAMs preserve an inherently explainable structure, making the model’s decisions more understandable and trustworthy. Finally, a formal approach for simplifying complex predictive models is proposed, based on representing models as sets of rules. This methodology enables users to control the trade-off between predictive accuracy and interpretability, thus providing a clear and actionable understanding of the factors responsible for failures. These contributions were validated on two complex industrial systems at the Bosch plant in Rodez: GRIBS, dedicated to the machining of diesel injector components, and RETCO HPC, used for deep drilling of high-pressure connectors. The results demonstrate that the proposed approaches enable reliable RUL prediction while offering meaningful interpretations of the underlying degradation mechanisms. This thesis thus highlights the crucial importance of integrating explainability and interpretability into predictive maintenance methods, in direct response to the operational needs of modern industrial systems.La maintenance prévisionnelle est devenue essentielle dans l’industrie 4.0, notamment pour les composants tournants critiques des systèmes complexes industriels, dont la défaillance engendre des coûts importants et des arrêts de production non planifiés. Cependant, la majorité des modèles prédictifs utilisés restent opaques et sont souvent perçus comme des "boîtes noires" par les opérateurs, limitant ainsi leur adoption sur le terrain. Cette thèse se concentre sur le développement d’approches interprétables et explicables en maintenance prévisionnelle, combinant performance prédictive et transparence pour améliorer leur acceptabilité. Pour atteindre cet objectif, plusieurs contributions complémentaires sont proposées. Tout d’abord, une méthodologie d’optimisation des fenêtres temporelles multivariées est développée pour améliorer la structuration des séries temporelles non stationnaires utilisées dans la maintenance prévisionnelle. Cette méthode permet d’automatiser la détermination des paramètres optimaux de segmentation temporelle, facilitant ainsi l’extraction d’informations pertinentes pour l’entraînement des modèles prédictifs. Les Multiclass Neural Additive Models (MNAM) sont ensuite introduits comme des modèles d’apprentissage automatique explicables, adaptés aux problématiques de classification multiclasses rencontrées en milieu industriel, telles que la prédiction de la durée de vie restante (RUL) des composants tournants, formulée sous forme de classes. Les MNAM préservent une structure explicable par construction, facilitant l’interprétation des décisions prises par le modèle. Enfin, une approche formelle de simplification de modèles prédictifs complexes est proposée. Celle-ci repose sur la représentation d’un modèle sous forme de règles, avec une méthodologie permettant de contrôler le compromis entre performance prédictive et interprétabilité, afin d’offrir aux utilisateurs une compréhension claire et exploitable des facteurs responsables des défaillances. Ces contributions ont été validées sur deux systèmes industriels complexes de l’usine Bosch à Rodez : GRIBS, destiné à l’usinage de composants pour injecteurs diesel, et RETCO HPC, utilisé pour le perçage profond de connecteurs haute pression. Les résultats montrent que les approches proposées permettent de prédire le RUL tout en offrant une interprétation des mécanismes de dégradation observés. Cette thèse démontre ainsi l’intérêt majeur d’intégrer explicabilité et interprétabilité aux méthodes prédictives en maintenance prévisionnelle, répondant directement aux besoins opérationnels des systèmes industriels modernes