Résumé : The rapid advances in speech synthesis and voice conversion technologies have fueled the rise of audio deepfakes, posing a serious threat to both the security of automatic speaker verification (ASV) systems and public trust in digital communication. Addressing these risks requires robust anti-spoofing countermeasures, supported by systematic exploration of datasets, model architectures, and training strategies.
This work examines the role of datasets and evaluation metrics in shaping progress, and analyzes the evolution of countermeasures under increasingly realistic conditions. The contributions focus on three main directions: (i) leveraging self-supervised learning (SSL) representations to capture both acoustic and linguistic cues; (ii) designing deep learning frameworks aimed at stronger generalizability; and (iii) exploring simple yet powerful back-ends in high-dimensional SSL spaces, which in some cases surpass more complex approaches by enhancing robustness and generalization.

Face Deepfake Detection: Overview and Open Challenges

Résumé : With the rapid progress of generative AI, the types of attacks on biometric systems have evolved from traditional spoofing attacks to deepfakes. While these techniques open creative opportunities, they also raise serious concerns related to misinformation, identity fraud, and digital security. This presentation explores the state of the art in facial deepfake detection and its intersection with face antispoofing. We will highlight their differences in threat models, methodologies, and application domains, as well as common challenges such as generalization, robustness, and domain adaptation. We will discuss emerging approaches such as multimodal fusion, and the potential of large language models (LLMs) to improve interpretability