MAYA: Multi-Attack Yielding Augmentation for Unified Face Attack Detection

초록

Face anti-spoofing remains a critical challenge as both physical and digital attack techniques continue to evolve. While prior methods often address either physical or digital threats in isolation, real-world attack scenarios require unified models capable of handling a broad range of spoofing attack strategies. In this work, we propose MAYA (Multi-Attack Yielding Augmentation), a unified face attack detection framework that generates subtype-specific pseudo-attacks during training. MAYA simulates eight distinct attack variants—including both physical (e.g., print, cutouts, replay) and digital (e.g., attribute-edit, face-swap, adversarial attack) threats—based on a fine-grained taxonomy of attack subtypes. To further enhance generalization, we introduce an attack-wise contrastive loss that enforces semantic consistency by aligning the features of samples sharing the same attack. Extensive experiments on the UniAttackData+ benchmark—the official dataset of the ICCV 2025 Face Anti-Spoofing Workshop Challenge—demonstrate the effectiveness of our method. Ablation studies further show that both the attack-wise contrastive loss and subtype-aware pseudo-attacks contribute to improved generalization across diverse attack types.