Invited Speakers

Abstract: Foundation models and multimodal LLMs have achieved significant performance on downstream tasks and promise zero-shot and few-shot generalisation in natural-language reasoning over images. In this talk, we focus on face biometrics and discuss if these models can be trusted in biometric applications. The talk explores where these models meet biometric recognition, attack detection, and security, and reports findings from systematic evaluations. Under standard verification protocols, multimodal LLMs capture useful semantic cues but trail specialised face recognition systems at high-precision operating points. For attack detection, in-context learning makes general-purpose models competitive, with the added benefit of explanations. We further investigate different approaches of adapting MLLMs for biometric security, which outperform state-of-the-art in the literature. The presentation is closed with open challenges and directions.