EPITA will host the Humanistica 2026 conference from May 19 to 22, 2026.

EPITA will be hosting the SSLR 2025 theme days from 08 to 09 October 2025.

This research aims to explore the use of generative AI, combined with explainability mechanisms, to improve proactive anomaly detection and the security of vehicle ad hoc networks (VANETs).

This seminar presents recent advances in automatic media analysis for gender representation, developed within the ANR-funded Gender Equality Monitor project. It reviews methods for measuring speech time, visual presence, and thematic content by gender, and discusses the integration of these indicators into public reports, along with the challenges of going beyond binary gender categories.

Two half-days dedicated to scientific presentations and discussions

This seminar presents recent advances in automatic media analysis for gender representation, developed within the ANR-funded Gender Equality Monitor project. It reviews methods for measuring speech time, visual presence, and thematic content by gender, and discusses the integration of these indicators into public reports, along with the challenges of going beyond binary gender categories.

This seminar presents a deep learning-based post-processing approach to improve the quality of breast tomosynthesis reconstructions, particularly in orthogonal slices where conventional methods struggle due to limited angular coverage. It also introduces a Bayesian framework to estimate both epistemic and aleatoric uncertainties, providing more reliable and interpretable reconstructed volumes.

The talk presents a conjectural characterisation of visibly pushdown languages (VPLs) in AC^0 obtained with Stefan Göller. The approach builds on recognisability by Ext-algebras, extending monoid morphism methods for regular languages. VPLs are classified into three categories: those in AC^0, those not in AC^0, and intermediate languages whose complexity remains unclear. The characterisation is complete up to the understanding of these intermediate languages.

This seminar presents a deep learning approach for the segmentation of guidewires in 2D fluoroscopic images during ERCP procedures, addressing challenges such as limited data, poor image quality, and the absence of dedicated methods. It also explores new evaluation strategies, both from a clinical and technical perspective, to better assess and improve segmentation performance.

Information extraction methods—such as named entity recognition, coreference resolution, or relation extraction—aim to produce structured data from unstructured or semi-structured sources like text. These structured outputs can then be used for various purposes, including database creation, ontology population, or automated reasoning. In the context of research in the humanities and social sciences (HSS), applying these methods is particularly valuable, especially when considering large-scale semantic annotation of corpora, thereby facilitating their critical study.

This project aims to create a multilingual catalog of popular literature by aggregating various data sources to link book content with reader engagement, thus enabling new perspectives on books as social and cultural artifacts.

In the context of Digital Humanities as a hybrid research field, we advocate for a genuinely interdisciplinary dialogue between the humanities and computer science—moving beyond mere tool application or engineering service—to promote the co-construction of methods, models, and interpretations, fostering a virtuous cycle grounded in epistemological collaboration, methodological reflexivity, and mutual disciplinary recognition.

In the context of Digital Humanities as a hybrid research field, we advocate for a genuinely interdisciplinary dialogue between the humanities and computer science—moving beyond mere tool application or engineering service—to promote the co-construction of methods, models, and interpretations, fostering a virtuous cycle grounded in epistemological collaboration, methodological reflexivity, and mutual disciplinary recognition.

The seminar presents a context-aware, graph-based anomaly detection framework for complex IoT ecosystems. This method models network traffic as a time-evolving multi-edge graph, capturing various interactions between devices. It uses scalable community detection and a Heterogeneous Graph Neural Network (HeteroGNN) to identify and classify anomalies in real-time, showing high accuracy and adaptability in experimental evaluations.

This talk presents a formal approach to constructing reliable and efficient composite fog services in distributed IoT environments. It combines functional, non-functional, and transactional aspects, using heuristic and adaptive composition strategies. The approach also includes a correct-by-construction verification method, illustrated on a distributed e-health application.

Active automata learning constructs models of software/hardware components from input/output observations. Timing information is crucial for many systems, motivating an extension of Mealy machine learning to machines with timers (MMTs). This algorithm generalizes the L# algorithm to timed settings using symbolic queries implemented via finitely many concrete queries.

Cette thèse explore l'inégalité algorithmique dans l'apprentissage automatique appliqué à l'éducation. En utilisant une nouvelle mesure d'impact des biais (MADD), l'étude révèle des inégalités non détectées par les méthodes existantes, affectant notamment les groupes historiquement marginalisés. Des solutions de réduction de ces biais et une librairie Python (maddlib) ont été développées pour faciliter la détection et la mitigation des discriminations algorithmiques en éducation.

Machine learning (ML) is revolutionizing content creation in game development, offering new possibilities for efficiency and innovation through generative models, diffusion techniques, and LLMs. Riot Games is exploring these technologies to assist with 3D asset creation, textures, animations, and interactive content, empowering artists rather than replacing them. This talk will delve into the challenges and opportunities of integrating AI into game development pipelines, highlighting best practices and emphasizing the importance of preserving creative control as these technologies evolve.

This half-day seminar, jointly organised by the TIRF team and the A3SI team from the LIGM laboratory presents recent advances on parallel algorithms for the construction of hierarchical reprensetations.

This study delves into how human experts detect junk codes, a known tool for evasion attacks. Through an online experiment with industry and academic practitioners, we analyze the detection of evasive samples by humans, from simple to complex tactics. We show that the experience in software development impacts the ability of the participants to identify junk codes correctly. Surprisingly, their experience in cyber security is not a great contribution to this task, although detecting malicious code is generally viewed as a field of cyber security. We also show that extended time slots do not aid in detecting such junk codes. Mostly, these types of malicious code are identified early, and additional time does not make a great improvement. We hope this study will emphasize the need for improved training for future experts, to enhance malware detection in human-computer defense systems. Also, as we show the detection time of skilled experts to be relatively short, we envision effective detection systems based on human-computer interaction in the future.

Les réseaux véhiculaires connaissent une interconnexion croissante, intégrant le réseau CAN pour le pilotage des véhicules, des systèmes d’infotainment, et le concept de “Software Defined Vehicle” (SDV) permettant la mise à jour et l’amélioration des capacités des véhicules après leur livraison. Si ces avancées offrent des opportunités considérables, notamment pour améliorer la sécurité, elles exposent également les véhicules à de nouvelles surfaces d’attaques potentielles.

In this talk, I will present recent work in my group that explores the use of advanced metadata encoding schemes to overcome limitations in memory safety overhead and checking accuracy.