Research map / Label-efficient and robust learning / Recognition and long-tail
Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HPDR, which learns leaf and non-leaf prototypes in hyperbolic (Poincaré ball) space with hierarchical prototype learning and a prototype-oriented classifier for generalizable face anti-spoofing.
Problem
Domain generalization for face anti-spoofing typically aligns source domains in a Euclidean feature space, which ignores the hierarchical relations among FAS samples (attack types, domain factors) and may hurt generalization to unseen domains and attacks.
Approach
Model FAS features in the Poincaré ball with learnable leaf prototypes (per class) and non-leaf prototypes forming a hierarchy; optimize prototype-instance, instance-instance, and prototype-prototype losses in hyperbolic space, and classify using hyperbolic distances to leaf prototypes via a learnable weighted head.
Key results
- On leave-one-out O&M&I to C, HPDR reaches 11.30% HTER and 94.42% AUC, the best among compared DG methods in Table 1.
- On limited source domains (M&I to C), HPDR achieves 22.22% HTER and 85.54% AUC, versus 24.52% HTER for AMEL.
- On HQ-WMCA cross-attack evaluation, HPDR gets 14.6±8.5 mean ACER versus 19.9±18.3 for DCN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HPDR instead of AMEL | DG face anti-spoofing · O&M&I to C | HTER (%): 11.88 -> 11.30 | Helped |
| use HPDR instead of SSAN-M | DG face anti-spoofing · O&C&M to I | HTER (%): 14.00 -> 11.26 | Helped |
| use HPDR instead of EBDG | DG face anti-spoofing · O&C&I to M | HTER (%): 9.56 -> 4.58 | Helped |
| use HPDR instead of AMEL | DG face anti-spoofing · I&C&M to O | HTER (%): 11.31 -> 9.93 | Helped |
| use HPDR instead of AMEL | limited-source DG face anti-spoofing · M&I to C | HTER (%): 24.52 -> 22.22 | Helped |
| use HPDR instead of AMEL | limited-source DG face anti-spoofing · M&I to O | HTER (%): 19.68 -> 21.07 | Hurt |
| use HPDR instead of Auxiliary(Depth) | cross-attack FAS · HQ-WMCA cross-attack mean | ACER (%): 22.3±18.2 -> 14.6±8.5 | Helped |
| add LPI, LII, and LPP losses (ablation) | HPDR · O&M&I to C | HTER (%): 17.78 -> 11.30 | Helped |
| add LPI, LII, and LPP losses (ablation) | HPDR · O&M&I to C | AUC (%): 90.29 -> 94.43 | Helped |
| add LPI only (ablation) | HPDR · O&M&I to C | HTER (%): 17.78 -> 15.56 | Helped |
| add LII only (ablation) | HPDR · O&M&I to C | HTER (%): 17.78 -> 15.64 | Helped |
| add LPP only (ablation) | HPDR · O&M&I to C | HTER (%): 17.78 -> 15.74 | Helped |
Surprising findings
- Using hyperbolic space alone without any of the proposed losses yields the worst result in the ablation, indicating the losses are needed for gains.
- Performance in the Euclidean case (c=0) degrades strongly (22.22% HTER), and HPDR performance drops for curvature values above 0.01.
Limitations
- Prototype count is bounded by batch size (30), so more prototypes cause unbalanced assignment and degrade results.
- Curvature sensitivity: performance peaks at c=0.01 and degrades at larger curvature because the exponential map at the origin becomes a poor approximation.
Experimental setup
Datasets: OULU-NPU CASIA-FASD MSU-MFSD Idiap Replay-Attack HQ-WMCA
Benchmarks: OULU-NPU CASIA-FASD MSU-MFSD Idiap Replay-Attack HQ-WMCA
Metrics: HTER AUC ACER
Baselines: MADDG SSDG-M DR-Net RFM D2AM ANRL DRDG SSAN-M AMEL EBDG CPL PCL
Builds on: DepthNet Poincaré ball model Dasgupta cost hierarchical clustering hyperbolic prototype learning
Closest papers by meaning
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.90)
- Gradient Alignment for Cross-Domain Face Anti-Spoofing (CVPR 2024 · similarity 0.89)
- Optimal Transport-Guided Source-Free Adaptation for Face Anti-Spoofing (CVPR 2025 · similarity 0.87)
- CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing (CVPR 2024 · similarity 0.87)
- From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.87)
- FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.