Content-adaptive face antispoofing for secure vision-based authentication

Face presentation attack detection (FacePAD) is essential for secure face authentication in mobile and embedded deployments, where computational budgets limit the use of heavy backbones, temporal stacks, or auxiliary sensors. We present a lightweight RGB-only FacePAD model that improves spatial selectivity by integrating content-adaptive spatial operators into a MobileNetV3 backbone. Instead of applying a fixed kernel everywhere, the proposed operator generates location-specific, channel-shared filters conditioned on the input feature content, following the involution principle. This design increases sensitivity to localized spoof artifacts (e.g., print noise, replay reflections, paper-mask edges) while maintaining mobile-class efficiency. Experiments on standard benchmarks—Replay-Attack, Replay-Mobile, and ROSE-Youtu—show strong performance, including zero error on Replay-Attack and Replay-Mobile and near-error rates on ROSE-Youtu. Overall, content-adaptive spatial filtering offers a practical route to robust FacePAD without extra modalities or temporal modeling.