Backbone and Fusion Method Adversarial Attack Vulnerability of Multibiometric Authentication

초록

Research on multibiometric authentication has been actively conducted to defend against adversarial attacks targeting specific biometric modalities. Multibiometric authentication systems select backbone models and fusion methods based on biometric types to extract patterns and context, and then integrate them as probability or feature values for user authentication. However, existing systems often select these components based solely on authentication performance, without evaluating vulnerability to adversarial attacks, which reduces system reliability. This paper proposes a vulnerability analysis system for adversarially robust multibiometric authentication, analysing the impact of backbone and fusion configurations. The system consists of a modelling stage, in which backbone models (VGGNet, ResNet, ViT, BEiT) and fusion methods (general, hierarchical, dense fusion) are selected, and an adversarial attack stage, in which adversarial noise is injected using FGSM and PGD. Experimental results show that for fingerprints, the ViT backbone with general fusion is most robust due to global pattern extraction and local context preservation, while for faces, the VGGNet backbone with dense fusion is most robust by extracting local patterns and preserving both local and global context. However, for adversarial attacks on faces, as noise intensity increases, the ViT backbone capable of global pattern extraction is more robust than the VGGNet backbone. In addition, the BEiT backbone, pretrained via masked image modelling that reconstructs object structure, is vulnerable to adversarial noise. Finally, the weighted mean method, which integrates extracted patterns and context as probability values for individual analysis, is more robust than concatenation, which integrates them as feature values for joint analysis. Fingerprints, with strong linear features, are also more robust than faces, which emphasise local object characteristics.

키워드