Toward Generalized and Explainable Face Anti-Spoofing via Decomposition-Fusion Networks

This paper presents DFNet, a decomposition–fusion network for generalized and interpretable face anti-spoofing. DFNet comprises two explicitly designed branches—a decomposition branch and a fusion branch—that jointly enhance robustness and interpretability in presentation attack detection. The decomposition branch extracts multi-level representations through a two-stage process, where coarse features and saliency maps are first computed to emphasize prominent regions, followed by the suppression of these regions to facilitate fine-grained feature extraction from less dominant areas. The fusion branch integrates the coarse and fine-grained streams using a residual architecture augmented with convolutional block attention modules, while a dynamic weighting mechanism adaptively balances their contributions during feature fusion. DFNet is extensively evaluated on multiple public benchmarks under both intra- and cross-dataset protocols, and its robustness is further assessed under strong Projected Gradient Descent (PGD) adversarial attacks. Experimental results demonstrate that DFNet achieves competitive or superior performance across diverse datasets and attack scenarios, while maintaining improved resilience to adversarial perturbations. Moreover, the modular, attention-driven design of DFNet provides saliency-based insights into the learned spoof traces, supporting its interpretability. Overall, this work high-lights the effectiveness of explicitly incorporating generalization, robustness, and interpretability into architectural design for real-world face anti-spoofing, with DFNet offering a concrete step toward this goal. For reproducibility, the source code is publicly available at https://github.com/dhimasarief/DFNet.