Adaptive RGB-Based Face Spoof Detection Using Multi-Expert Feature Decomposition and Gated Fusion

Face spoofing attacks—including print, replay, and digitally manipulated facial presentations—pose a significant threat tomodern biometric authentication systems. Conventional spoof detection approaches often rely on single-stream convolutionalmodels or specialized hardware sensors such as depth or infrared cameras, limiting their robustness and deployability in realworld consumer devices. To address these challenges, this paper proposes an adaptive multi-expert spoof detection framework that operates exclusively on a single RGB input while achieving enhanced generalization against both known and unseen attacks. The architecture decomposes spoof detection into three complementary expert branches: an RGB appearance expert, a depth-aware structural expert, and a frequency-domain expert. To effectively integrate these heterogeneous cues, a reliability-aware gated fusion mechanism is introduced to dynamically weight expert contributions. Furthermore, the framework incorporates an adaptive feedback mechanism that continuously refines the model by leveraging high-confidence samples during deployment. Extensive experiments demonstrate that the proposed approach achieves superior performance across diverse attack scenarios compared to state-of-the-art methods.