Research map / Label-efficient and robust learning / Domain adaptation and generalization
Gradient Alignment for Cross-Domain Face Anti-Spoofing
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 GAC-FAS, a sharpness-aware objective for domain generalization in face anti-spoofing that aligns domain-wise SAM ascending-point gradients with the ERM gradient to converge to flat minima, without extra auxiliary modules.
Problem
Domain generalization for face anti-spoofing: detectors trained on source datasets generalize poorly to unseen domains, and prior DG methods do not ensure convergence to flat minima.
Approach
Extends SAM to multi-domain FAS by computing per-domain ascending points and regularizing the generalization gradient at these points to align with the ERM gradient over all source domains, so the update is robust to domain shift.
Key results
- Leave-one-out ICM→O HTER 8.60% and OCM→I HTER 4.29%, establishing state-of-the-art over prior DG-FAS methods in Table 1.
- Limited source domains MI→C HTER reduced from 24.07% to 16.91%.
- Upon convergence, OMI→C HTER reduced from 19.40% to 15.37%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GAC-FAS instead of UDG-FAS | Cross-domain FAS · OCM → I | HTER: 5.86 -> 4.29 | Helped |
| Use GAC-FAS instead of IADG | Cross-domain FAS · OCI → M | HTER: 5.41 -> 5.00 | Helped |
| Use GAC-FAS instead of IADG | Cross-domain FAS · OMI → C | HTER: 8.70 -> 8.20 | Helped |
| Use GAC-FAS instead of the strongest listed baseline | Cross-domain FAS · OCI → M | AUC: 98.75 -> 97.56 | Hurt |
| Use GAC-FAS instead of IADG | Limited-source FAS · MI → C | HTER: 24.07 -> 16.91 | Helped |
| Use GAC-FAS instead of IADG | Limited-source FAS · MI → O | AUC: 90.49 -> 89.67 | Hurt |
| Use GAC-FAS instead of UDG-FAS | Unseen 2D attacks · Unseen 2D attack | AUC: +0.96% AUC | Helped |
| Use GAC-FAS instead of UDG-FAS | Unseen 3D attacks · Unseen 3D attack | AUC: +2.01% AUC | Helped |
| Replace GAC-FAS with the gradient regularizer ⟨∇Lpi(Si), ∇L(S)⟩ (ablation) | GAC-FAS · ICM → O | HTER: 9.89 -> 11.35 | Hurt |
| Use domain-wise gradients with SAM (ablation) | SAM · ICM → O | HTER: 11.99 -> 12.51 | Hurt |
| Use domain-wise gradients with ASAM (ablation) | ASAM · ICM → O | HTER: 11.73 -> 12.18 | Hurt |
| Use domain-wise gradients with SAGM (ablation) | SAGM · ICM → O | HTER: 11.63 -> 12.19 | Hurt |
Surprising findings
- Many SAM variants applied to domain-wise gradients did not surpass SoTA DG-FAS baselines, with domain-wise gradients prone to noisy ascending-point estimates.
Limitations
- Hyperparameters gamma and rho were not tuned per task and higher values deteriorate performance.
- Performance with gamma set to zero is only competitive with SoTA rather than best.
- Focus limited to domain shift; robustness to other noise types not studied.
Experimental setup
Datasets: Idiap Replay Attack OULU-NPU CASIA-MFSD MSU-MFSD CelebA-Spoof
Benchmarks: Idiap Replay Attack OULU-NPU CASIA-MFSD MSU-MFSD CelebA-Spoof
Metrics: HTER AUC TPR95
Baselines: SSDG-M SSDG-R SSAN IADG SA-FAS UDG-FAS PathNet AMEL SAM ASAM GSAM SAGM
Builds on: Sharpness-Aware Minimization SAGM GSAM Fish SA-FAS
Closest papers by meaning
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.92)
- CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing (CVPR 2024 · similarity 0.89)
- Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space (CVPR 2024 · similarity 0.89)
- Domain-Hallucinated Updating for Multi-Domain Face Anti-spoofing (AAAI 2024 · similarity 0.89)
- From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.87)
- SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.87)
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.