Reproduce “NAS-FAS: Static-Dynamic Central Difference Network Search for Face Anti-Spoofing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zitong Yu, Jun Wan, Yunxiao Qin, Xiaobai Li, Stan Z. Li, Guoying Zhao
NAS-FAS searches for face anti-spoofing architectures that combine central-difference operations, static-dynamic visual features, depth supervision, and domain- or attack-type-aware optimization. It reports ACER values as low as 0.2% on OULU-NPU and 0.04% on SiW, while cross-dataset evaluation shows stronger generalization to unseen attacks, including 85.78% AUC and 15.00% HTER on the challenging CASIA-SURF 3DMask dataset.
Key insights
- Central-difference operators provide a useful inductive bias for detecting local spoofing artifacts; the strongest reported setting uses difference weights of 0.7 for both convolution and pooling.
- Motion information helps when combined with appearance, but dynamic-only inputs lose important spatial cues; the proposed fused representation improves over static RGB by about 0.3 percentage points in ACER.
- Architecture search quality depends strongly on the FAS-specific search space and supervision: generic NAS with cross-entropy alone performs substantially worse than the task-aware configuration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.