Research map / Label-efficient and robust learning / Domain adaptation and generalization
CFPL-FAS: Class Free Prompt Learning for Generalizable 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 CFPL, a CLIP-based class-free prompt learning framework that generates content and style prompts via Q-Formers, uses prompt-text matching, diversified style mixing, and prompt modulation to improve domain-generalized face anti-spoofing.
Problem
Domain generalization for face anti-spoofing: existing methods rely on domain labels or instance-level disentanglement, which distorts semantic feature structure and limits generalization to unseen domains.
Approach
Uses CLIP's text encoder as a source of modulation factors rather than classifier weights: Content and Style Q-Formers produce class-free prompts from visual content and style features, and the resulting text features gate channel-wise visual features via a prompt modulation module, with a prompt-text matching loss and style-statistics mixing for diversification.
Key results
- On Protocol 1 leave-one-out, CFPL achieves 3.60% average HTER, versus 7.82% for the previous best (IADG) without CelebA-Spoof.
- With CelebA-Spoof as supplementary data, CFPL* reaches 2.98% average HTER on Protocol 1, slightly better than FLIP-MCL* (3.01%).
- On Protocol 2, CFPL achieves 10.88% average HTER versus 13.87% for CLIP; CFPL* reaches 7.01% average HTER.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CFPL instead of IADG | DG FAS · Protocol 1 average | HTER: 7.82 -> 3.60 | Helped |
| Use CFPL instead of CLIP-V | DG FAS · CASIA-SURF (CW→S) | HTER: 8.32 -> 8.77 | Hurt |
| Add prompt-text matched (PTM) supervision to the prompt-learning baseline (ablation) | CFPL · Protocol 2 average on all sub-protocols | HTER: 8.11 -> 7.50 | Helped |
| Add diversified style prompt (DSP) mixing (ablation) | CFPL · Protocol 2 average on all sub-protocols | HTER: 7.50 -> 7.08 | Helped |
| Add prompt modulation (PM) with the designed gate function (ablation) | CFPL · Protocol 2 average on all sub-protocols | HTER: 7.08 -> 6.72 | Helped |
| Use both CQF and SQF instead of either one alone (ablation) | CFPL · ICM→O | HTER: 5.12 -> 3.33; 4.84 -> 3.33 | Helped |
| Increase query length from 16 to 32 with depth 1 (ablation) | CFPL · ICM→O | HTER: 3.33 -> 3.33 | No clear effect |
| Increase Q-Former depth from 1 to 8 with query length 16 (ablation) | CFPL · ICM→O | HTER: 3.33 -> 3.56 | Hurt |
Surprising findings
- On CASIA-SURF in Protocol 2, the plain CLIP-V baseline outperforms CFPL on HTER, AUC and TPR, which the authors attribute to obvious spoofing traces making visual features alone relatively generalizable.
- Removing the gate function and replacing it with a simple mean of content and style text features reduces performance only modestly (HTER 6.72% to 7.08%).
Limitations
- Gains on CASIA-SURF are absent; CFPL underperforms CLIP-V on that target domain in Protocol 2.
- Relies on CelebA-Spoof as supplementary data to reach the best reported numbers, and some comparisons use 5-shot target data.
Experimental setup
Datasets: MSU-MFSD CASIA-FASD Replay-Attack OULU-NPU CASIA-SURF CASIA-SURF CeFA WMCA CelebA-Spoof
Benchmarks: MSU-MFSD CASIA-FASD Replay-Attack OULU-NPU CASIA-SURF CASIA-SURF CeFA WMCA
Metrics: HTER AUC TPR@FPR=1%
Baselines: MADDG DR-MD-Net RFMeta NAS-FAS D2AM SDA DRDG ANRL SSDG-R SSAN-R PatchNet SA-FAS
Builds on: CLIP CoOp BLIP-2 MixStyle ALBEF
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)
- Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models (AAAI 2025 · similarity 0.87)
- From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.87)
- Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space (CVPR 2024 · similarity 0.87)
- FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · 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.