Publications

2025

• Deep Learning for Embedded Cybersecurity
  
  • Varillon Arnaud
  , 2025. Public-key cryptography is one of the core pillars of cybersecurity, in particular thanks to the authentication schemes it enables. It is embedded in many ubiquitous objects, such as hardware wallets. Side-channel attacks are one of the major threats to such devices. In particular, the advances in machine learning, and more specifically deep learning, which have marked the last ten years, seem likely to make such attacks remarkably effective. In such a hostile environment, assessing the true security level of devices intended for cryptographic use is of utmost importance: indeed, it has now even become vital to the smooth running of information systems.In this thesis, we have appraised the security, of implementations, reputed to be the more secure ones in the face of such attacks (“power” and “EM” channels), which manipulate the secret key bit by bit. Numerous contributions, which mainly use deep learning, have been published on this subject. Unfortunately, none of them provides any guarantee regarding the robustness of the device under study in the face of such attacks: each time, the method being described does not allow one to state with certainty that it is not possible to find a more powerful attack than that being presented. The security level of the latter devices is therefore potentially underestimated. More specifically, fundamental aspects of the classification task associated with any attack, such as the shape of its decision boundary or the optimality - in terms of attack performance - of the features derived from the sampling of the side-channel under consideration, are never addressed. Therefore, we have sought to find methodologies that are as close to optimality as possible given the conditions imposed by the exercise (for example, the possibility of configuring the key used by the target for in-depth analyses).Initially, assuming that an attacker can control the key that is parameterized in the target, only vertical attacks are considered. In this context, the optimal effectiveness of the joint use of NICV, for feature selection, and the perceptron, for classification, is highlighted from a theoretical point of view. In particular, the security of a cryptographic library hitherto considered robust (libecc) is called into question. Secondly, assuming that an attacker cannot set the key as he wishes in the target, yet still has a functionally perfect clone, another procedure is proposed for carrying out the security evaluation, this time using horizontal (collision-based) attacks based on an unsupervised learning technique which, because it requires (by definition) minimal training at most, is better suited to such a scenario. Compared to the state of the art, the approach followed is closer to optimality without however achieving it, but avenues are suggested to get there in the near future. Last, to validate these findings, experimental verifications have been carried out on a board (STM32F407) which features a Cortex-M4 processor that can be found in many hardware wallets (e.g. Trezor Model T).
• Planning Socio-Emotional Response Generation for Conversational Agents
  
  • Vanel Lorraine
  , 2025. Conversational systems are now capable of producing impressive and generally relevant responses. However, we have no visibility or control over the socio-emotional strategies behind state-of-the-art large language models (LLMs). This poses a problem in terms of their transparency and thus their trustworthiness for critical applications, such as industrial customer service virtual agents. This thesis aims to develop socially intelligent conversational systems that can model the conversational dynamics of user and agent social behaviours. To train models in the role of customer service agents, we need to capture the various nuances of their speech patterns, which often combine emotional support and problem-solving skills. We design a multi-label approach, modelling the behaviours expressed in one speaker as a chronologically ordered sequence of socio-emotional labels, including emotions and conversational actions.We propose a two-module architecture. The planning module uses the dialogue's history to generate the best sequence of social and emotional strategies for the response. The generative module conditions the response generation according to the predicted sequence of labels to improve the quality of the final answer while allowing for transparency and controllability. To efficiently adapt the system to our use case, we require adequate datasets; however, no publicly available corpus meets all our criteria, including multi-label annotations and French task-oriented conversations. To address this, we introduce DATA-SERGE, a new corpus composed of French customer service conversations, annotated following a comprehensive multi-label scheme. Additionally, we observe that existing automatic metrics are lacking to properly evaluate such a system. We present a human evaluation protocol and new scores to fill this gap.We evaluate this architecture on both DailyDialog, an English open-domain dataset, and DATA-SERGe. The findings support our primary research hypothesis and demonstrate that, even in highly domain-specific contexts like French customer service data, planning sequences of labels has a positive impact on response generation.
• RESCUE: Multi-Robot Planning Under Resource Uncertainty and Objective Criticality
  
  • Cordeiro Franco
  • Tardieu Samuel
  • Pautet Laurent
  , 2025, 37th Euromicro Conference on Real-Time Systems (ECRTS 2025), pp.5:1-5:23. Robot planning in distributed systems, such as drone fleets performing active perception missions, presents complex challenges. These missions require cooperation to achieve objectives like collecting sensor data or capturing images. Multi-robot systems offer significant advantages, including faster execution and increased robustness, as robots can compensate for individual failures. However, resource costs, affected by environmental factors such as wind or terrain, are highly uncertain, impacting battery consumption and overall performance. Mission objectives are often prioritized by criticality, such as retrieving data from low-battery sensors to prevent data loss. Addressing these priorities requires sophisticated scheduling to navigate high-dimensional state-action spaces. While heuristics are useful for approximating solutions, few approaches extend to multi-robot systems or adequately address cost uncertainty and criticality, particularly during replanning. The Mixed-Criticality (MC) paradigm, extensively studied in real-time scheduling, provides a framework for handling cost uncertainty by ensuring the completion of high-critical tasks. Despite its potential, the application of MC in distributed systems remains limited. To address the decision-making challenges faced by distributed robots operating under cost uncertainty and objective criticality, we propose four contributions: a comprehensive model integrating criticality, uncertainty, and robustness; distributed synchronization and replanning mechanisms; the incorporation of mixed-criticality principles into multi-robot systems; and enhanced resilience against robot failures. We evaluated our solution, named RESCUE, in a simulated scenario and show how it increases the robustness by reducing the oversizing of the system and completing up to 40% more objectives. We found an increase in resilience of the multi-robot system as our solution not only guaranteed the safe return of every non-faulty robot, but also reduced the effects of a faulty robot by up to 14%. We also computed the performance gain compared to using MCTS in a single robot of up to 2.31 for 5 robots. (10.4230/LIPIcs.ECRTS.2025.5)
  DOI : 10.4230/LIPIcs.ECRTS.2025.5
• Side-Channel Attack Detection using gem5 and Machine Learning: A Case Study on Fault-based Attacks in RISC-V
  
  • Khan Mahreen
  • Mushtaq Maria
  • Pacalet Renaud
  • Apvrille Ludovic
  , 2025. <div><p>Microarchitectural side-channel attacks pose a significant threat to modern computing architectures. This paper presents a machine learning-based methodology for detecting these attacks using the gem5 simulator, focusing on the recently discovered Flush+Fault attack [6] on RISC-V. Our approach follows a three-phase process. The first phase is data collection, where we simulate attack and non-attack scenarios in gem5 and extract microarchitectural features indicative of side-channel activity. The second phase is the training phase, where we utilize machine learning (ML) techniques to build a classification model capable of distinguishing between normal execution and attack patterns. The last phase is the testing phase, where we evaluate the trained model using various performance metrics to validate its accuracy and precision. To the best of our knowledge, this is the first detection framework for Flush+Fault attacks [6] on RISC-V, showcasing its effectiveness in mitigating emerging threats. Our results indicate that gem5 metrics combined with machine learning models can reliably detect Flush+Fault attacks, achieving 0.99 accuracy with random forest (RF), 0.96 with support vector machine (SVM), and 0.95 with naïve bayes (NB). Moreover, this methodology is adaptable to different side-channel attacks and architectures, making it a promising approach for strengthening microarchitectural security.</p></div>
• Side-Channel Attack Detection using gem5 and ML: A Case Study on Fault-based Attacks in RISC-V
  
  • Khan Mahreen
  • Mushtaq Maria
  • Pacalet Renaud
  • Apvrille Ludovic
  , 2025. <div><p>Microarchitectural timing side-channel attacks exploit variations in execution times caused by the underlying hardware to extract sensitive information. These attacks leverage architectural features like caches and branch predictors. For thorough analysis, we use Gem5 simulations to analyze Flush+Fault attack behavior on RISC-V.</p></div>
• Neural networks for solition prediction in optical communication
  
  • Mukhopadhyay Meghna
  • Thedrez Bruno
  , 2025. <div><p>We study a neural network approach for predicting the number of solitons in WDM signals, extending it to the computation of the nonlinear Fourier transform eigenvalues in the case of Satsuma-Yajima signals.</p></div>
• Quantum-aware network planning and integration
  
  • Ware Cédric
  • Lourdiane Mounia
  , 2025. In order to broaden the adoption of highly-demanded quantum functionalities such as QKD, there is a need for having quantum signals coexist with classical traffic over the same physical medium, typically optical fibers in already-deployed networks. Beyond the experimental point-to-point demonstrations of the past few years, efforts are now underway to integrate QKD at the network level: developing interfaces with the software-defined-network ecosystem; but also network planning tools that satisfy physical-layer contraints jointly on the classical and quantum signals. We have found that in certain situations, naïve network planning prioritizing quantum traffic drastically degrades classical capacity, whereas a quantum-aware wavelength assignment heuristic allows coexistence with minimal impact on both capacities. More such techniques will be required to enable widespread deployment of QKD and other future quantum functionalities. (10.1109/ICTON67126.2025.11125476)
  DOI : 10.1109/ICTON67126.2025.11125476
• First Order Logic with Fuzzy Semantics for Describing and Recognizing Nerves in Medical Images
  
  • Bloch Isabelle
  • Bonnot Enzo
  • Gori Pietro
  • La Barbera Giammarco
  • Sarnacki Sabine
  , 2025. This article deals with the description and recognition of fiber bundles, in particular nerves, in medical images, based on the anatomical description of the fiber trajectories. To this end, we propose a logical formalization of this anatomical knowledge. The intrinsically imprecise description of nerves, as found in anatomical textbooks, leads us to propose fuzzy semantics combined with first-order logic. We define a language representing spatial entities, relations between these entities and quantifiers. A formula in this language is then a formalization of the natural language description. The semantics are given by fuzzy representations in a concrete domain and satisfaction degrees of relations. Based on this formalization, a spatial reasoning algorithm is proposed for segmentation and recognition of nerves from anatomical and diffusion magnetic resonance images, which is illustrated on pelvic nerves in pediatric imaging, enabling surgeons to plan surgery.
• Stein's method for max-stable random vectors
  
  • Costacèque Bruno
  • Decreusefond Laurent
  , 2025. Motivated by the omnipresence of extreme value distributions in limit theorems involving extremes of random processes, we adapt Stein's method to include these laws as possible target distributions. We do so by using the generator approach of Stein's method, which is possible thanks to a recently introduced family of semi-groups. We study the corresponding Stein solution and its properties when the working distance is either the smooth Wasserstein distance or the Kolmogorov distance. We make use of those results to bound the distance between two max-stable random vectors, as well as to get a rate of convergence for the de Haan-LePage series in smooth Wasserstein distance. (10.48550/arXiv.2507.00463)
  DOI : 10.48550/arXiv.2507.00463
• Arithmétisation de la partie entière et applications à la cryptographie
  
  • Berthet Pierre-Augustin
  • Tavernier Cédric
  , 2025. Dans ce travail nous proposons une méthodologie pour calculer la partie entière à l'aide uniquement d'additions et de multiplications, la rendant ainsi compatible avec le chiffrement homomorphe.
• Cartographier l'intelligence artificielle : collecte et analyse d'un corpus de chartes et manifestes sur l'éthique de l'IA
  
  • Viard Tiphaine
  • Delarue Simon
  • Gornet Mélanie
  • Boritchev Maria
  , 2025. Nous présentons ici un corpus de 436 chartes et manifestes au sujet de l'éthique de l'intelligence artificielle, ainsi qu'une analyse discursive de celui-ci, permettant de comprendre les enjeux autour desquels le monde social de l'éthique de l'IA se structure. Nous mettons à disposition le corpus, sa documentation, et le code permettant son analyse.
• Rapid Thermalization of Dissipative Many-Body Dynamics of Commuting Hamiltonians
  
  • Kochanowski Jan
  • Alhambra Álvaro
  • Capel Ángela
  • Rouzé Cambyse
  Communications in Mathematical Physics, Springer Verlag, 2025, 406 (8), pp.176. Abstract Quantum systems typically reach thermal equilibrium rather quickly when coupled to a thermal environment. The usual way of bounding the speed of this process is by estimating the spectral gap of the dissipative generator. However the gap, by itself, does not always yield a reasonable estimate for the thermalization time in many-body systems: without further structure, a uniform lower bound on it only constrains the thermalization time to grow polynomially with system size. Here, instead, we show that for a large class of geometrically-2-local models of Davies generators with commuting Hamiltonians, the thermalization time is much shorter than one would naïvely estimate from the gap: at most logarithmic in the system size. This yields the so-called rapid mixing of dissipative dynamics. The result is particularly relevant for 1D systems, for which we prove rapid thermalization with a system size independent decay rate only from a positive gap in the generator. We also prove that systems in hypercubic lattices of any dimension, and exponential graphs, such as trees, have rapid mixing at high enough temperatures. We do this by introducing a novel notion of clustering which we call “strong local indistinguishability” based on a max-relative entropy, and then proving that it implies a lower bound on the modified logarithmic Sobolev inequality (MLSI) for nearest neighbour commuting models. This has consequences for the rate of thermalization towards Gibbs states, and also for their relevant Wasserstein distances and transportation cost inequalities. Along the way, we show that several measures of decay of correlations on Gibbs states of commuting Hamiltonians are equivalent, a result of independent interest. At the technical level, we also show a direct relation between properties of Davies and Schmidt dynamics, that allows to transfer results of thermalization between both. (10.1007/s00220-025-05353-y)
  DOI : 10.1007/s00220-025-05353-y
• Clustering cell nuclei on microgrooves for disease diagnosis using deep learning
  
  • Roellinger Bettina
  • Thenier Francois
  • Leclech Claire
  • Coirault Catherine
  • Angelini Elsa
  • Barakat Abdul I
  Scientific Reports, Nature Publishing Group, 2025, 15 (1), pp.22476 (1-13). Various diseases including laminopathies and certain types of cancer are associated with abnormal nuclear mechanical properties that influence cellular and nuclear deformations in complex environments. Recently, microgroove substrates designed to mimic the anisotropic topography of the basement membrane have been shown to induce 3D nuclear deformations in various adherent cell types. Importantly, these deformations are different in myoblasts derived from laminopathy patients from those in cells derived from normal individuals. Here we assess the ability of a Variational Autoencoder (VAE) and a Gaussian Mixture Model (GMM) to cluster patches of nuclei of both wildtype myoblasts and myoblasts with laminopathy-associated mutations cultured on microgroove substrates, and we explore the impact of image processing parameters on clustering performance. We show that a standard VAE with GMM is able to cluster nuclei based on their morphologies and degrees of deformations and that these clusters correspond to either wildtype myoblasts or myoblasts with LMNA mutations. The current results suggest that combining deep learning techniques with microgroove substrates enables automatic classification of nuclear deformations and thus provides a promising approach for easy and rapid diagnosis of pathologies that involve abnormalities in nuclear deformation. (10.1038/s41598-025-05788-2)
  DOI : 10.1038/s41598-025-05788-2
• Versatile Rectifiers 2.0: Theory, Design, and Emerging Applications
  
  • Niotaki Kyriaki
  • Song Chaoyun
  • Yang Furong
  • Georgiadis Apostolos
  IEEE Microwave Magazine, Institute of Electrical and Electronics Engineers, 2025, 26 (7), pp.93-102. Conventional rectifiers are used primarily for the rectification of ac, RF, or microwave power into dc. Their major applications are concentrated in energy harvesting (EH) and wireless power transfer (WPT) [1], [2], [3], [4], [5]. Rectifiers can be used as stand-alone components or in combination with complementary energy sources, such as solar or wind. They can also be used beyond simple power conversion. Backscattering and simultaneous wireless information and power transmission (SWIPT) systems are two particularly interesting application examples [6]. (10.1109/MMM.2024.3470708)
  DOI : 10.1109/MMM.2024.3470708
• Neurosymbolic Methods for Dynamic Knowledge Graphs
  
  • Alam Mehwish
  • Gesese Genet Asefa
  • Paris Pierre-Henri
  , 2025. Knowledge graphs (KGs) have recently been used for many tools and applications, making them rich resources in structured format. However, in the real world, KGs grow due to the additions of new knowledge in the form of entities and relations, making these KGs dynamic. This chapter formally defines several types of dynamic KGs and summarizes how these KGs can be represented. Additionally, many neurosymbolic methods have been proposed for learning representations over static KGs for several tasks such as KG completion and entity alignment. This chapter further focuses on neurosymbolic methods for dynamic KGs with or without temporal information. More specifically, it provides an insight into neurosymbolic methods for dynamic (temporal or non-temporal) KG completion and entity alignment tasks. It further discusses the challenges of current approaches and provides some future directions. (10.3233/FAIA250223)
  DOI : 10.3233/FAIA250223
• Neurosymbolic Methods for Rule Mining
  
  • Ławrynowicz Agnieszka
  • Galárraga Luis
  • Alam Mehwish
  • Jaulmes Bérénice
  • Zeman Václav
  • Kliegr Tomáš
  , 2025, pp.1-22. In this chapter, we address the problem of rule mining, beginning with essential background information, including measures of rule quality. We then explore various rule mining methodologies, categorized into three groups: inductive logic programming, path sampling and generalization, and linear programming. Following this, we delve into neurosymbolic methods, covering topics such as the integration of deep learning with rules, the use of embeddings for rule learning, and the application of large language models in rule learning. (10.3233/FAIA250224)
  DOI : 10.3233/FAIA250224
• Prototyping Custom Hardware Performance Counters in gem5 Simulator: A Framework for RISC-V Side-Channel Attack Assessment
  
  • Khan Mahreen
  • Mushtaq Maria
  • Pacalet Renaud
  • Apvrille Ludovic
  , 2025. <div><p>Microarchitectural side-channel attacks exploit performancecentric optimizations such as caches and branch predictors to leak sensitive data. While RISC-V processors are increasingly adopted, tools to evaluate their resilience to these attacks remain limited. This paper presents the first framework for designing custom hardware performance counters (HPCs) in the gem5 simulator for RISC-V, targeting the assessment of side-channel attacks. Although gem5 itself can provide valuable statistics to assess attacks, real-world systems do not have access to such detailed insights. To address this gap, we extend gem5 by creating a new component to prototype HPCs for RISC-V. We modify gem5 to create three HPCs to evaluate branch mispredictions, and L1 instruction and data cache misses. These parameters are critical for identifying microarchitectural side-channel attack patterns. To validate our framework, we test it under various workload conditions on Flush+Fault [7], a recent RISC-V attack. Our simulations detect a 491.4% increase in branch mispredictions and a 231.4% rise in instruction cache misses during Flush+Fault execution compared to baseline benign OS processes. These deviations directly align with its exploitation mechanisms of branch mistraining and cache interference. The framework correlates these HPC deviations with attack attempts, enabling vulnerability analysis of components such as cache and branch predictors. Furthermore, it lays the foundation for a prototype evaluation platform that integrates custom HPCs into gem5 for testing realistic detection mechanisms and countermeasures. To our knowledge, no prior work implements custom HPCs in gem5 for RISC-V security analysis. Our goal is to establish a RISC-V security evaluation platform for analyzing attack patterns, testing detection mechanisms, and guiding secure hardware design.</p></div>
• Neural Velocity for hyperparameter tuning
  
  • Dalmasso Gianluca
  • Bragagnolo Andrea
  • Tartaglione Enzo
  • Fiandrotti Attilio
  • Grangetto Marco
  , 2025, pp.1-8. Hyperparameter tuning, such as learning rate decay and defining a stopping criterion, often relies on monitoring the validation loss. This paper presents NeVe, a dynamic training approach that adjusts the learning rate and defines the stop criterion based on the novel notion of "neural velocity". The neural velocity measures the rate of change of each neuron’s transfer function and is an indicator of model convergence: sampling neural velocity can be performed even by forwarding noise in the network, reducing the need for a held-out dataset. Our findings show the potential of neural velocity as a key metric for optimizing neural network training efficiently. (10.1109/IJCNN64981.2025.11229080)
  DOI : 10.1109/IJCNN64981.2025.11229080
• Denoising Diffusion Probabilistic Model for Point Cloud Compression at Low Bit-Rates
  
  • Spadaro Gabriele
  • Presta Alberto
  • Giraldo Zuluaga Jhony Heriberto
  • Grangetto Marco
  • Hu Wei
  • Valenzise Giuseppe
  • Fiandrotti Attilio
  • Tartaglione Enzo
  , 2025. Efficient compression of low-bit-rate point clouds is critical for bandwidth-constrained applications. However, existing techniques mainly focus on high-fidelity reconstruction, requiring many bits for compression. This paper proposes a "Denoising Diffusion Probabilistic Model" (DDPM) architecture for point cloud compression (DDPM-PCC) at low bit-rates. A PointNet encoder produces the condition vector for the generation, which is then quantized via a learnable vector quantizer. This configuration allows to achieve a low bitrates while preserving quality. Experiments on ShapeNet and ModelNet40 show improved rate-distortion at low rates compared to standardized and state-of-the-art approaches. We publicly released the code at https://github.com/EIDOSLAB/DDPM-PCC.
• Décoder le pouvoir de persuasion dans les concours d'éloquence : une étude sur la capacité des modèles de langues à évaluer la prise de parole en public
  
  • Barkar Alisa
  • Chollet Mathieu
  • Labeau Matthieu
  • Biancardi Beatrice
  • Clavel Chloé
  , 2025, pp.77-90. L’importance des compétences en prise de parole en public (PPP) stimule le développement de systèmes d’évaluation automatisée, mais l’intégration des grandes modèles de langue (LLMs) reste peu explorée. Nous proposons un cadre où les LLMs évaluent des critères issus de la littérature et de retours de formateurs. Nous testons trois approches : des prédictions LLM directes à zéro coup (RMSE 0, 8) par rapport à des prédictions de persuasion basées sur des caractéristiques lexicales fabriquées à la main (RMSE 0, 51) ou basées sur des critères évalués par LLM 0, 6 insérés en entrée dans ElasticNet. L’analyse des liens entre critères et caractéristiques lexicales montre que seul le critère de niveau de langue évalué par LLM est prévisible (score F1 de 0, 56) soulignant les limites actuelles des LLMs pour l’analyse de la PPP. Code source et données disponibles sur GitHub.
• Power-Aware digital twin of coherent optical receiver
  
  • Purkayastha Ambashri
  • Delezoide Camille
  • Boitier Fabien
  • Bajaj Vinod
  • Lourdiane Mounia
  • Ware Cédric
  • Layec Patricia
  , 2025. <div><p>We propose a digital twin of coherent receivers based on an extended physical model for predicting quality of transmission under receiver power variations. We experimentally validate it, demonstrating an accuracy improved by up to 1.5dB.</p></div>
• Deep reinforcement learning for wind farm flow control
  
  • Kadoche Elie
  , 2025. Within wind farms, wake interactions between turbines can significantly reduce overall energy production. Wind farm flow control encompasses methods designed to mitigate these effects through coordinated turbine control. Wake steering, for example, involves redirecting the wake of certain machines to optimize airflow and increase power output. This is achieved via yaw control, by intentionally misaligning certain turbines with the incoming wind. However, designing robust wake steering controllers is challenging: they must adapt to dynamic and uncertain wind conditions, respect turbine actuation constraints, and capture the complexity of physical interactions in large arrays. Reinforcement learning offers a powerful framework for developing more effective controllers. By leveraging principles from artificial intelligence, it enables wind farms to adapt and respond intelligently to diverse wind conditions. A reinforcement learning agent learns from past experience and continuously refines control strategies as the wind changes. This Ph.D. thesis further studies and develops deep reinforcement learning-based wake steering controllers for wind farm flow control. More specifically: 1) it studies the impact of wind dynamics on yaw control and highlights the importance of optimizing strategies over a certain time horizon when variations in wind direction are significant; 2) it investigates how multi-agent reinforcement learning approaches can control large-scale wind farms and capture complex wake interactions in time-varying wind conditions; 3) it leverages self-attention mechanisms within a single-agent setting to build more capable controllers, demonstrating significant improvements in sample efficiency, learning performance and generalization. Overall, this Ph.D. thesis contributes to narrowing the gap between algorithmic advances in deep reinforcement learning and their practical deployment for large-scale wind farm flow control.
• Bias in Face Recognition : Assessment and Post-Processing Mitigation
  
  • Conti Jean-Rémy
  , 2025. Face Recognition (FR) systems have reached unprecedented accuracy due to deep learning advances, enabling widespread deployment. Yet, concerns about fairness—especially across demographic subgroups defined by sensitive attributes like gender and ethnicity— persist. This thesis addresses algorithmic bias in FR through two complementary lenses: evaluation and mitigation.The first part introduces a statistically rigorous framework for assessing FR performance and fairness. Focusing on similarity-based ROC analysis, we apply U-statistics theory to derive asymptotically valid confidence intervals for key metrics. We also propose a scalar uncertainty score to quantify the statistical reliability of fairness assessments. This enables practitioners to identify disparities that are not only large but statistically significant.The second part tackles post-processing bias mitigation for frozen, pre-trained FR models. We begin with gender bias, proposing a novel mitigation approach using two operationally relevant fairness metrics: BFAR and BFRR, which capture disparities in false acceptance/rejection rates under fixed-security constraints. We introduce the Ethical Module, a shallow neural network that transforms embeddings using a Fair von Mises-Fisher loss. This loss models group-specific intra-class variance, enabling fairness control without access to sensitive attributes at deployment. The method improves fairness with minimal performance loss and low computational cost.Next, we address racial bias via a post-processing strategy aligned with FR objectives. Using centroid-based pseudo-scores as proxies for similarity, we develop the Centroid Fairness loss, implemented through a lightweight Fairness Module. This module, trained on top of a frozen model, facilitates subgroup alignment and effectively mitigates racial disparities while preserving overall accuracy—challenging the conventional fairness-performance trade-off.Together, these contributions offer a statistically grounded and practically effective toolbox for evaluating and mitigating bias in face recognition. By integrating uncertainty quantification, fairness-aware representation learning, and scalable post-hoc debiasing, this thesis is an additional step towards the development of equitable and trustworthy biometric technologies.
• Efficient Thermalization and Universal Quantum Computing with Quantum Gibbs Samplers
  
  • Rouzé Cambyse
  • França Daniel Stilck
  • Alhambra Álvaro
  , 2025, pp.1488-1495. The preparation of quantum Gibbs states is a crucial task in quantum computing. In this work, we prove that a recently introduced, efficiently implementable dissipative evolution thermalizes to the Gibbs state in time scaling polynomially or even logarithmically with system size at high enough temperatures for any Hamiltonian that satisfies a Lieb-Robinson bound, such as local Hamiltonians on a lattice. Furthermore, we show the efficient adiabatic preparation of the associated purifications or "thermofield double" states. To the best of our knowledge, these are the first results rigorously establishing the efficient preparation of high-temperature Gibbs states and their purifications. In the low-temperature regime, we show that implementing this family of dissipative evolutions for inverse temperatures polynomial in the system's size is computationally equivalent to standard quantum computations. On a technical level, for high temperatures, our proof makes use of the mapping of the generator of the evolution into a Hamiltonian, and then connecting its convergence to that of the infinite temperature limit. We further present an alternative proof that is based on showing the exponential decay of the so-called oscillator norm, yielding convergence in logarithmic times. For low temperature, we instead perform a perturbation at zero temperature and resort to circuit-to-Hamiltonian mappings akin to the proof of universality of quantum adiabatic computing. Taken together, our results show that a family of quasi-local dissipative evolutions efficiently prepares a large class of quantum many-body states of interest, and has the potential to mirror the success of classical Monte Carlo methods for quantum many-body systems. (10.1145/3717823.3718268)
  DOI : 10.1145/3717823.3718268
• Home Spaces and Semiflows for the Analysis of Parameterized Petri Nets
  
  • Memmi Gerard
  , 2025, Vol-3998, pp.97-113. After briefly recalling basic notations of Petri Nets, home spaces, and semiflows, we focus on ℱ + , the set of semiflows with non-negative coordinates where the notions of minimality of semiflows and minimality of supports are particularly critical to develop an effective analysis of invariants and behavioral properties of Petri Nets such as boundedness or even liveness. We recall known behavioral properties attached to the notion of semiflows that we associate with extremums. We also recall three known decomposition theorems considering N, Q + , and Q respectively where the decomposition over N is being improved with a necessary and sufficient condition.<p>Then, we regroup a number of properties (old and new) especially around the notions of home spaces and home states which, in combination with semiflows, are used to efficiently support the analysis of behavioral properties. We introduce a new result on the decidability of liveness under the existence of a home state. We introduce new results on the structure and behavioral properties of Petri Nets, illustrating again the importance of considering generating sets of semiflows with non-negative coordinates.</p><p>As examples, we present two related Petri Net modeling arithmetic operations (one of which represents an Euclidean division), illustrating how results on semiflows and home spaces can be methodically used in analyzing the liveness of the parameterized model and underlining the efficiency brought by the combination of these results to the verification engineer.</p>