Réunion

Impersonation Attacks on Biometric Recognition Systems

Oratrice : Marta Gomez-Barrero, Universität der Bundeswehr München

Résumé

With the widespread use of biometric recognition, several issues related to the privacy and security provided by this technology have been raised and analysed. Starting with the classification of attack points published by Ratha et al. in 2001, recent works have analysed the vulnerability of biometric systems from different perspectives, including inverse biometrics or morphing attacks. The former constitutes a severe threat for biometric systems from two different angles: sensitive personal data (i.e., biometric data) can be derived from compromised unprotected templates, and other powerful attacks can be launched building upon synthetic reconstructed samples. Morphing attacks, on the other hand, can severely decrease the security of identity documents or processes, in which morphed images matching two or more identities are injected. The present talk will describe these attack forms under the umbrella term of impersonation attacks, and discuss their impact on biometric technologies, including standardised evaluation metrics.

A Story of ZK Proofs: How Efficient ZK-Proofs Enable Signatures and New Applications

Oratrice : Emmanuela Orsini, Bocconi University

Résumé

Zero-Knowledge (ZK) proofs have evolved from theoretical constructs into powerful tools that drive practical innovations in cryptography and security. 

In this talk, we explore the remarkable development of ZK proofs, focusing on prover-efficient schemes based on VOLE (Vector Oblivious Linear Evaluation), which drastically improve proof efficiency and scalability for large statements. We illustrate how these efficient ZK proofs enable advanced cryptographic primitives, particularly post-quantum signatures. Beyond signatures, we also highlight novel applications emerging in privacy-preserving authentication, secure computation and machine learning.

Demystifying Modern Scams; Breaking the Stigma and Building Resilience

Oratrice : Leyla Bilge, Gen

Résumé

As online scams evolve in scale and sophistication, traditional defenses are no longer sufficient to protect users from financial and emotional harm. At Gen Digital's Research Labs, we are pioneering a new approach to digital safety—one that combines deep cybersecurity expertise with domain-trained AI systems capable of detecting fraud before it reaches the user. In this talk, I will present the latest advancements in our AI-powered anti-scam technologies, including SMS and email protection, real-time browser and messaging app defenses, and our most recent AI assistant for scam prevention. I will share key findings from our research on scammer behavior, highlight regional trends in AI-enabled scams, and discuss the growing role of AI tools on this tooic.

TBA

Oratrice : Ksenia Ermoshina, CNRS

Résumé

Evolving Digital Forensic Sciences

Oratrice invitée : Lena Klasén, Swedish Police Authority, Stockholm, Linköping University, Sweden

Résumé

Forensic science faces significant challenges due to rapid societal and technological changes, requiring adaptation to increasing data volume, integration complexities, and evolving ethical concerns. The field is transitioning from traditional methodologies to a more digital-centric approach, incorporating data from IoT devices, smartphones, and a multitude of diverse sources. This shift blurs the distinction between physical and virtual traces, challenging how we ensure contextual accuracy, quality, and relevance in forensic work.

Digital forensics plays a crucial role in this transformation, leveraging AI and advanced technologies to convert digital traces into actionable evidence. While AI enhances forensic capabilities, it also presents risks, such as algorithmic bias, data privacy concerns, and not the least the misuse of emerging technologies by criminals. The forensic community must stay ahead by integrating AI responsibly, ensuring efficiency, robustness, and reliability in forensic investigations.

Future directions include fostering a shared forensic science culture, embracing digital transformation for forensic intelligence and interdisciplinary collaboration. Emerging technologies, such as multimodal AI for 3D crime scene modeling (digital twins), are revolutionizing evidence analysis by enabling large-scale monitoring, rapid data processing, and virtual reconstructions of crime scenes. This talk explores these advancements, highlighting both opportunities and challenges in digital forensics and forensic science.

Obscuration réversible de chiffres dans des images

Orateur : Valentin Noyé, LIRMM

Résumé

Dans un contexte où la protection de la vie privée est de mise, notamment avec la prolifération d’images contenant des données sensibles, nous nous intéressons à l’obscuration d’informations visuelles, en particulier des chiffres pouvant être contenus dans des images (documents administratifs, plaques d'immatriculations, ...). Dans nos travaux, nous proposons une méthode fondée sur la manipulation du vecteur latent issu d’un auto-encodeur afin d’obscurcir les valeurs des chiffres. L'obscuration non visible d'une image consiste à protéger son contenu en le transformant en un autre contenu cohérent. Notre méthode permet d'obscurcir les chiffres présents dans une image tout en préservant une haute fidélité visuelle et en garantissant la réversibilité grâce à une clé secrète. La qualité de la transformation est évaluée à l’aide de critères à la fois visuels et quantitatifs. De plus, nous introduisons des mécanismes d'attaque et de défense afin d'évaluer la robustesse de notre méthode.

PAD datasets: Constraints and hidden factors in their conception

Orateur : Peter Marsh, GREYC

Résumé

Datasets are crucial for training deep learning models, as AI is increasingly being utilised for security purposes such as Presentation Attack Detection (PAD). However, significant constraints are often overlooked during the creation of these datasets. In this presentation, I will discuss the key challenges and limitations encountered while developing the dataset for the FACEPAD3D project, led by the GREYC.

Désapprentissage pour la Mitigation d’Attaques en Apprentissage et le Droit à l’Oubli sans Réentraînement Coûteux

Orateur : Achraf Kerzazi, Orange Labs

Résumé

Tandis que l’IA est de plus en plus utilisée dans l’industrie pour des tâches critiques, la confiance accordée à ces systèmes exige une analyse approfondie de leurs vulnérabilités. Avec l’émergence de menaces sans précédent, des attaques sur les données jusqu’à celles du modèle lui-même, il devient évident qu’il est nécessaire de mettre en place des mécanismes de défense solides. Dans cet exposé, nous abordons une question de recherche centrale : comment éliminer les effets d’une attaque sur un modèle d’apprentissage profond sans nécessiter le réentraînement de celui-ci à partir de zéro ? Cela devient de plus en plus critique avec l’explosion du volume de données, et le droit à l’oubli, que ce soit légal ou éthique, signale également la nécessité de telles mesures. Notre travail fait partie de l’effort pour envisager et mettre en œuvre des solutions en la matière. Ainsi, nous cherchons à développer une méthode efficace basée sur le désapprentissage, ou machine unlearning, pour garantir la sécurité et la conformité tout en préservant les performances des systèmes d’apprentissage concernés.

Leveraging Image Quality for Face PAD: IQA-Aware PAD

Orateur : Hamza Bouzid, GREYC

Résumé

Several studies have shown that presentation attacks (PAs) often introduce subtle but detectable degradations in image quality, such as unnatural textures, lighting inconsistencies, or printing artifacts. These cues make Image Quality Assessment (IQA) a promising direction for improving Presentation Attack Detection (PAD) systems. This work explores the effect of including IQA in face PAD, starting with investigating the correlation between natural scene statistics (NSS)-based IQA features and PAD performance, and evaluating models that integrate explicit IQA information. While still in its early stages, this study goal is to better understand the role of image quality in face PAD and guide future improvements.