Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Domain Generalization 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 Test-Time Domain Generalization (TTDG) for face anti-spoofing, projecting unseen test samples' style statistics onto learnable style bases (TTSP) trained with diversity and content-consistency losses (DSSS), with no test-time model updates.
Problem
Domain generalization for face anti-spoofing degrades when unseen target domains differ largely from source domains; existing DG methods only learn domain-invariant features at train time and do not exploit test data.
Approach
Use unlabeled test samples at inference by projecting their style statistics (channel-wise mean/std) onto a set of learnable style bases via cosine-similarity softmax weighting and AdaIN, with style-diversity and content-consistency losses in a hyperspherical feature space to learn the bases; no model parameters are updated at test time.
Key results
- Leave-one-out I&C&M to O: TTDG reaches 10.00% HTER / 95.70% AUC with CNN backbone, vs 11.45% / 94.50% for IADG.
- O&C&M to I: TTDG-V (ViT-Base) reaches 9.62% HTER / 98.18% AUC; TTDG CNN reaches 6.50% HTER / 97.98% AUC.
- Limited source domains (M&I to C): TTDG achieves 17.77% HTER vs 23.51% for IADG, outperforming prior DG methods by 5-6% HTER.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply TTDG instead of the strongest CNN baseline | FAS cross-domain model · I&C&M to O | HTER (%): 11.45 -> 10.00 | Helped |
| apply TTDG instead of the strongest CNN baseline | FAS cross-domain model · O&C&M to I | HTER (%): 11.04 -> 6.50 | Helped |
| apply TTDG instead of the strongest CNN baseline | FAS cross-domain model · O&C&I to M | HTER (%): 8.45 -> 7.91 | Helped |
| apply TTDG instead of the strongest CNN baseline | FAS cross-domain model · O&M&I to C | HTER (%): 11.88 -> 8.14 | Helped |
| add TTSP style projection (ablation) | TTDG · I&C&M to O | HTER (%): +3.40 | Helped |
| add both style diversity and content consistency losses (ablation) | TTDG · I&C&M to O | HTER (%): 13.64 -> 10.00 | Helped |
| add both style diversity and content consistency losses (ablation) | TTDG · O&C&M to I | HTER (%): 13.75 -> 6.50 | Helped |
| add both style diversity and content consistency losses (ablation) | TTDG · O&C&I to M | HTER (%): 13.33 -> 7.91 | Helped |
| add both style diversity and content consistency losses (ablation) | TTDG · O&M&I to C | HTER (%): 16.48 -> 8.14 | Helped |
| use content consistency loss without style diversity loss (ablation) | TTDG · I&C&M to O | HTER (%): 13.64 -> 11.63 | Helped |
| use style diversity loss without content consistency loss (ablation) | TTDG · I&C&M to O | HTER (%): 13.64 -> 13.05 | Helped |
Surprising findings
- Learnable style bases without the DSSS losses give only modest gains, and the style-diversity loss alone yields limited improvement, suggesting both losses are needed together.
Limitations
- Relies on a large number of style bases (N=64) and hyperparameter tuning (lambda_c=0.4).
- Gains reported only on the four standard FAS datasets with leave-one-out and limited-source protocols.
- Test-time projection requires test batches of unseen data to be available at inference.
Experimental setup
Datasets: CASIA-MFSD Idiap Replay-Attack MSU-MFSD OULU-NPU
Benchmarks: CASIA-MFSD Idiap Replay-Attack MSU-MFSD OULU-NPU
Metrics: HTER AUC
Baselines: MADDG D2AM SSDG RFM DRDG ANRL FGHV SSAN AMEL EBDG IADG ViTranZFAS
Builds on: AdaIN MADDG IADG Test-time style shifting (TTSS) FPS sampling
Closest papers by meaning
- Gradient Alignment for Cross-Domain Face Anti-Spoofing (CVPR 2024 · similarity 0.92)
- CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing (CVPR 2024 · similarity 0.90)
- Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space (CVPR 2024 · similarity 0.90)
- Domain-Hallucinated Updating for Multi-Domain Face Anti-spoofing (AAAI 2024 · similarity 0.88)
- SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.88)
- From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.88)
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.