Mixture-of-Attack-Experts with Class Regularization for Unified Physical-Digital Face Attack Detection

• DOI: 10.1609/aaai.v39i2.32218
• Publication Date: 2025-04-11T09:35:16Z
• Authors: Shunxin Chen; Ajian Liu; Junze Zheng; Jun Wan; Kailai Peng; Sergio Escalera; Zhen Lei
• Keywords: Regularization

Unified detection of digital and physical attacks in facial recognition systems has become a focal point research recent years. However, current multi-modal methods typically ignore the intra-class inter-class variability across different types attacks, leading to degraded performance. To address this limitation, we propose MoAE-CR, framework that effectively leverages class-aware information for improved attack detection. Our improvements manifest at two levels, i.e., feature loss level. At level, Mixture-of-Attack-Experts (MoAEs) capture more subtle differences among various fake faces. introduce Class Regularization (CR) through Disentanglement Module (DM) Cluster Distillation (CDM). The DM enhances class separability by increasing distance between centers live face classes. center-to-center constraints alone are insufficient ensure distinctive representations individual features. Thus, CDM further cluster features around their while maintaining separation from other Moreover, specific significantly deviate common patterns often overlooked. issue, our calculation prioritizes distant Extensive experiments on unified physical-digital datasets demonstrate state-of-the-art performance proposed method.