Archive ouverte de Centrale Lyon
Not a member yet
32420 research outputs found
Sort by
Exploiting Nearest Neighbor Mixing: Nearest Neighbor Poisoning
No full textFederated Learning (FL) enables multiple data holders (workers) to collaboratively train models without exposing their local datasets. However, some participants may behave in faulty or malicious ways (Byzantine), sending corrupted updates that hinder convergence. Byzantine FL seeks to ensure the robustness of distributed learning algorithms against such adversarial behavior. The challenge becomes more severe under data heterogeneity, where distinguishing Byzantine machines from benign but atypical ones is difficult. The Nearest Neighbor Mixing (NNM) defense has been proposed as a theoretically optimal solution, claiming strong robustness guarantees and demonstrating impressive empirical results.In this paper, we challenge these guarantees by introducing Nearest Neighbor Poisoning (NNP), a novel attack specifically designed to break NNM. We empirically show that NNP substantially degrades model performance across diverse heterogeneous settings-sometimes causing more than a 50% higher test error compared to other stateof-the-art attacks. Even against defenses other than NNM, NNP often remains the most damaging attack. These results highlight that robust Byzantine FL under heterogeneous data remains an open and unresolved problem. Our implementation is publicly available 1 .</div
PriviRec: Confidential and Decentralized Graph Filtering for Recommender Systems
No full textInternational audienceRecent advances in recommender systems have shown that relying on graph filters, such as the normalized item-item adjacency matrix and the ideal low-pass filter yields competitive performance and scales better than Graph Convolutional Networks-based solutions. However, these solutions require centralizing user data, which raises concerns over data privacy, security, and the monopolization of user data by a few actors. To address those concerns, we propose PriviRec and PriviRec-k, two complementary recommendation frameworks. In PriviRec, we show that it is possible to decompose widely used filters so that they can be computed in a distributed setting using Secure Aggregation and a distributed version of the Randomized Power Method, without revealing individual users contributions. PriviRec-k extends this approach by having users securely aggregate low-rank projections of their contributions, enabling a tunable balance between communication overhead and recommendation accuracy. We demonstrate theoretically as well as experimentally on Gowalla, Yelp2018, and Amazon-Book that our methods achieve performance comparable to centralized state-of-the-art recommender systems and superior to decentralized ones, while preserving confidentiality and low communication and computational overheads
Accelerating Signed Distance Functions
No full textInternational audienceProcessing and particularly visualizing implicit surfaces remains computationally intensive when dealing with complex objects built from construction trees. We introduce optimization nodes to reduce the computational cost of the field function evaluation for hierarchical construction trees, while preserving the Lipschitz or conservative properties of the function. Our goal is to propose acceleration nodes directly embedded in the construction tree, and avoid external, accompanying data structures such as octrees. We present proxy and continuous level of detail nodes to reduce the overall evaluation cost, along with a normal warping technique that enhances surface details with negligible computational overhead. Our approach is compatible with existing algorithms that aim at reducing the number of function calls. We validate our methods by computing timings as well as the average cost for traversing the tree and evaluating the signed distance field at a given point in space. Our method speeds up signed distance field evaluation by up to three orders of magnitude, and applies to both ray-surface intersection computation in Sphere Tracing applications and polygonization algorithms
Représentation condensée de RDF et son application dans le versionnement de graphe
No full textEvolving phenomena, often complex, can be represented using knowledge graphs, which have the capability to model heterogeneous data from multiple sources. Nowadays, a considerable amount of sources delivering periodic updates to knowledge graphs in various domains is openly available. The evolution of data is of interest to knowledge graph management systems, and therefore it is crucial to organize these constantly evolving data to make them easily accessible and exploitable for analysis. In this article, we will present and formalize the condensed representation of these evolving graphs and propose a new solution called QuaQue that allows querying across multiple versions of graphs and we also present the results of our benchmark comparing our solution against existing approaches.Les phénomènes évolutifs, souvent complexes, peuvent être représentés à l'aide de graphes de connaissances, qui ont la capacité de modéliser des données hétérogènes provenant de multiples sources. De nos jours, un nombre considérable de sources fournissant des mises à jour périodiques aux graphes de connaissances dans divers domaines sont librement accessibles. L'évolution des données présente un intérêt pour les systèmes de gestion des graphes de connaissances. Il est donc essentiel d'organiser ces données en constante évolution afin de les rendre facilement accessibles et exploitables à des fins d'analyse. Dans cet article, nous présenterons et formaliserons la représentation condensée de ces graphes évolutifs et proposerons une nouvelle solution appelée QuaQue qui permet d'effectuer des requêtes sur plusieurs versions de graphes. Nous présenterons également les résultats de notre benchmark comparant notre solution aux approches existantes
A Cascade of Mesostrophy in Turbulence with Reduced Vortex-Stretching
No full textIn three-dimensional turbulence, vortex stretching is the central mechanism enabling the transfer of kinetic energy toward smaller scales. If vortex stretching is absent in three-dimensional turbulence, energy is no longer conserved and enstrophy cascades to smaller scales. In this paper we propose a system which interpolates between these two cases by reducing the strength of vortex stretching in the Navier-Stokes equations. The resulting dynamics yield a cascade that is neither a pure energy cascade nor a pure enstrophy cascade. We refer to this process as a mesostrophy cascade. We formulate a consistent picture describing this cascade and the characteristic scales involved in the mesostrophy balance. We illustrate this picture via numerical integrations of the EDQNM model with reduced vortex stretching
Estimation de la longueur de cohérence en envergure d'un cylindre en écoulement à partir du champ lointain
No full textInternational audienceL’influence de la forme d’un cylindre en écoulement sur la longueur de cohérence en envergure de son sillage est étudiée. Si l’écoulement est tridimensionnel, les fluctuations de forces ne sont pas en phase sur l’envergure mouillée totale du cylindre, mais seulement sur quelques diamètres. Cette cohérence en envergure a déjà été mesurée au fil chaud dans la littérature, et il a été montré qu’elle augmente avec le rapport d’aspect du cylindre. Or, elle est constitutive de l’efficacité des dipôles rayonnés. Ainsi, la forme du cylindre altère son efficacité acoustique. Le but est de retrouver ces résultats à l’aide d’une méthode non-invasive, c’est-à-dire grâce à une antenne de microphones placée en champ lointain. L’idée est d’abord de simuler des sources plus ou moins cohérentes le long d’une ligne pour vérifier l’efficacité de la méthode dans la mesure de la longueur de cohérence, puis de placer des cylindres de formes diverses dans une soufflerie anéchoïque. Ainsi, on peut montrer que les résultats de mesure du champ de vitesse en écoulement concordent avec celles du champ de pression mesuré avec une antenne de microphones
Espaces de Sobolev à valeurs variétés
No full textThis thesis is concerned with several properties of Sobolev spaces of mappings between manifolds. An important amount of research has been carried out on these spaces since the beginning of the 80's, notably motivated by their strong connection with problems arising from geometry, from physics, or from numerical methods. Although they are defined as metric subspaces of classical Sobolev spaces of vector-valued mappings, spaces of mappings with values into a manifold exhibit striking qualitative differences with the former ones. A typical instance of such a difference is the fact that smooth maps into a given manifold need not be dense among the Sobolev mappings with values into the same target, in strong contrast with the classical density result in real-valued functions spaces. Following this observation, a whole area of research was initiated, focused notably on the four following questions: (I) characterize those values of the parameters s and p, the domains, and the targets for which strong density of smooth maps does occur; (II) find a suitable class of almost smooth maps which is always dense among Sobolev mappings; (III) when strong density fails, characterize the closure of smooth maps; and (IV) determine what happens if strong convergence is replaced by a weaker notion. This thesis aims at presenting a contribution in the study of each of these questions. First, we solve the missing case s > 1 noninteger in the first two questions, where the main difficulty is the combination of the rigidity of higher order spaces with the nonlocal character in fractional order, hence concluding the complete answer to the strong density problem. We also push further the study of the second question by establishing the strong density of an improved class of almost smooth maps. Then, we construct two families of analytical obstructions to the weak approximation property, showing that, for any p in N {0,1} — the only case to be left open — there exist targets so that the weak approximation property fails. Finally, we construct integral invariants allowing to characterize the closure of smooth maps for a large class of target manifolds in the range 0 1 non entier des deux premières questions, où la difficulté principale est la combinaison de la rigidité des espaces d'ordre supérieur et du caractère non local propre à l'ordre fractionnaire, finalisant ainsi la réponse complète au problème de la densité forte. Par ailleurs, on pousse plus avant l'étude de la deuxième question en établissant la densité forte d'une classe améliorée d'applications presque lisses. Ensuite, on construit deux familles d'obstructions analytiques à la propriété d'approximation faible, montrant que pour tout p dans N {0,1} — le seul cas encore ouvert — il existe des cibles pour lesquelles la propriété d'approximation faible échoue. Enfin, on construit des invariants intégraux permettant de caractériser la clôture des fonctions lisses pour une grande famille de variétés cibles dans la gamme 0 < s < 1, correspondant au cas où la construction même de tels invariants est une tâche délicate
Enhancing Dynamic Control of Inertial District Energy Networks Through a Physics-Informed State-Space Model
No full textInternational audienc
Online Health Monitoring of Silicon PV Panels by Converter-Based Impedance Spectroscopy: Panel-Level Equivalent Circuit Model and Health Feature Extraction
No full textInternational audienceWithin the framework of Dexter’s theory, we calculate the energies of the Stark levels of Yb3+-Yb3+ paired centers in lithium niobate doped with Yb3+ ions (LiNbO3:Yb3+) crystal, considering the interaction of optical electrons of ytterbium ions forming the paired center. The calculated Stark level energies are shown to correspond well with the observed cooperative luminescence wavelengths
SysTemp : Système multi-agents par la génération basée sur un templatede squelettes pour le SysML v2
No full textInternational audienceLe SysML v2 est un langage de modélisation pour l'ingénieurie de systèmes à base de modèles (MBSE). La génération de modèles SysML v2 à l'aide de l'intelligence artificielle représente un défi majeur dans l'ingénierie des systèmes complexes, notamment en raison de la rareté des corpus d'apprentissage. Nous présentons une contribution qui vise à faciliter et à améliorer la création de modèles SysML v2 à partir de spécifications en langage naturel. Nous proposons une approche basée sur un système multi-agents, avec introduction d'un générateur de templates, qui permet de guider la génération et introduire de l'information. Nous discutons des avantages et des défis de ce système à travers une série d'évaluations, soulignant son potentiel pour améliorer la productivité et la qualité dans la modélisation SysML v2.</div