Research map / Label-efficient and robust learning / Domain adaptation and generalization
Optimal Transport-Guided Source-Free Adaptation for 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 OTA, a source-free few-shot test-time adaptation method for face anti-spoofing that keeps a CLIP ViT-B/16 backbone frozen and adapts via optimal-transport-shifted prototypes (training-free) or a linear classifier trained with geodesic mixup (lightweight).
Problem
Face anti-spoofing models generalize poorly to client-specific target domains and new attacks, while clients cannot share source data or host model parameters and have only a few labeled target samples.
Approach
Learn multi-centroid prototypes as source-domain summaries, then adapt at test time by optimal transport of these prototypes toward few-shot target features (training-free) or by training a lightweight classifier on geodesic mixup samples interpolated along the Wasserstein geodesic between source prototypes and target features.
Key results
- Lightweight OTA reaches average HTER 2.91% across the four cross-domain protocols, versus 3.60% for CFPL (the best prior DG baseline in Table 1).
- Under unseen 3D attack protocol, OTA lightweight achieves AUC 99.63 versus 89.27 for GAC-FAS.
- Training-free OTA achieves 3.21% average HTER with zero learnable parameters and about 0.17 minutes of adaptation time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use lightweight test-time adaptation with geodesic mixup instead of training-free adaptation | OTA · Cross-domain, Avg. | HTER: 3.21 -> 2.91 | Helped |
| Use lightweight adaptation instead of zero-shot inference | OTA · Cross-domain, Avg. | HTER: 3.43 -> 2.91 | Helped |
| Use geodesic mixup rather than Manifold Mixup | OTA lightweight adaptation · Cross-domain, Avg. | HTER: 3.04 -> 2.91 | Helped |
| Add orthogonal loss to prototype loss (ablation) | OTA prototype framework · Cross-attack, 3D | HTER: 5.20 -> 5.14 | Helped |
| Add coarse supervised contrastive loss (ablation) | OTA prototype framework · Cross-attack, 3D | HTER: 5.14 -> 4.87 | Helped |
| Add fine-grained supervised contrastive loss (ablation) | OTA prototype framework · Cross-attack, 3D | HTER: 4.87 -> 4.19 | Helped |
| Increase sub-centroids from 1 to 50 per class (ablation) | OTA prototype framework · Cross-attack, 3D | HTER: 9.40 -> 5.20 | Helped |
| Increase sub-centroids from 30 to 50 per class (ablation) | OTA prototype framework · Cross-attack, 3D | HTER: 5.07 -> 5.20 | No clear effect |
| Adapt with one-class target examples rather than zero-shot (ablation) | OTA training-free · Cross-domain, MSU-MFSD (OCI→M) | HTER: 2.62 -> 2.38 | Helped |
| Adapt with one-class examples rather than both-class examples (ablation) | OTA training-free · Cross-domain, CASIA-MFSD (OMI→C) | HTER: 2.22 -> 2.33; 2.22 -> 2.67 | Mixed |
Surprising findings
- Providing few-shot data from the bona fide class alone is more effective than from the spoof class in most cross-domain scenarios.
Limitations
- Paper's main contribution is on par with DG baselines without test-time adaptation; gains depend on few-shot target data being available.
- Lightweight adaptation requires about 22.7 minutes, which is slower than the training-free variant.
- Evaluation uses only standard FAS benchmarks with a single backbone (CLIP ViT-B/16).
Experimental setup
Datasets: OULU-NPU CASIA-MFSD Idiap Replay Attack MSU-MFSD CelebA-Spoof
Benchmarks: OULU-NPU CASIA-MFSD Idiap Replay Attack MSU-MFSD CelebA-Spoof
Metrics: HTER AUC TPR@FPR=1%
Baselines: SSDG-R SSAN-R SA-FAS IADG CFPL ViTAF NCM Tip-Adaptor Linear Probe Manifold Mixup
Builds on: ArcFace Mixup Sinkhorn algorithm Tip-Adaptor SupCon
Closest papers by meaning
- Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space (CVPR 2024 · similarity 0.87)
- Gradient Alignment for Cross-Domain Face Anti-Spoofing (CVPR 2024 · similarity 0.87)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.86)
- FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.85)
- One-Class Face Anti-spoofing via Spoof Cue Map-Guided Feature Learning (CVPR 2024 · similarity 0.85)
- From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.85)
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.