da/sec scientific talk on Biometrics

Topic: In-context Learning Inference of Vision Language Models for Physical and Digital Attacks

by Lazaro Janier Gonzalez-Soler
D19/2.03a, April 17, 2025 (Thursday), 12.00 noon

Keywords — In-context learning, presentation attack detection, morphing attack detection, vision language models

Abstract

„Recent research in biometric systems has made significant strides in addressing fraudulent practices. However, as detection methods improve, attack techniques become increasingly sophisticated. Attacks on face recognition systems can be broadly divided into physical and digital approaches. Traditionally, deep learning models have been the primary defense against such attacks. While these models perform exceptionally well in scenarios for which they have been trained for, they often struggle to adapt to different types of attacks or varying environmental conditions. These subsystems require substantial amounts of training data to achieve reliable performance, yet biometric data collection faces significant challenges including privacy concerns and the logistical difficulties of capturing diverse attack scenarios under controlled conditions. This work investigates the application of Vision Language Models (VLM) for detecting physical presentation attacks and digital morphing attacks in biometric systems. Focusing on open-source models under 8 billion parameters, the first systematic framework for quantitative evaluation of VLMs in security-critical scenarios through in-context learning techniques is established. The experimental evaluation conducted on freely available databases demonstrate that VLMs achieve competitive performance for physical and digital attack detection, outperforming some of the traditional CNNs without resource exhaustive training, and imply insightful trajectory for differential morphing attack detection challenge. The results validate VLMs as promising tools for improving generalization in attack detection.“