Research map / Vision-language and multimodal / CLIP and zero-shot
SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining
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 SLIP, a CLIP-based one-class face anti-spoofing framework that uses language-guided spoof cue map estimation, prompt-driven liveness feature disentanglement, and spoof-like feature augmentation to learn liveness from live images only.
Problem
One-class face anti-spoofing trained only on live faces struggles to separate live from spoof and tends to entangle liveness features with live/spoof-irrelevant domain information (e.g., facial content), hurting cross-domain generalization.
Approach
Uses CLIP prompts describing live, spoof (object-covering), content and hybrid concepts, paired with zero or nonzero pseudo spoof cue maps, to supervise a one-class model; disentangles liveness from content via contrastive prompt losses and augments spoof-like features by fusing live image features with spoof prompt features.
Key results
- On OULU-NPU intra-domain protocol 1, SLIP reaches 14.58% ACER versus 23.49% for OC-SCMNet (CVPR 24).
- On cross-domain OCM→I, SLIP achieves 15.71% HTER and 89.38% AUC, the best among one-class methods.
- Averaged across the protocols, SLIP improves ACER by about 14.79% over prior one-class FAS methods on OULU-NPU intra-domain testing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Gaussian-noise pseudo-spoof features with prompt learning for live and spoof features (ablation) | one-class FAS baseline · OCM→I | HTER (%): 34.43 -> 26.64 | Helped |
| add prompt-driven liveness feature disentanglement ($\mathcal{L}_{FD}$) (ablation) | SLIP · OCM→I | HTER (%): 26.64 -> 23.93 | Helped |
| add live-image feature alignment ($\mathcal{L}_{FA}$) (ablation) | SLIP · OCM→I | HTER (%): 26.64 -> 22.43 | Helped |
| add both $\mathcal{L}_{FD}$ and $\mathcal{L}_{FA}$ (ablation) | SLIP · OCM→I | HTER (%): 26.64 -> 19.50 | Helped |
| add spoof-like image feature augmentation ($\mathcal{L}_{A}$) (ablation) | SLIP · OCM→I | HTER (%): 19.50 -> 15.71 | Helped |
| include position adjectives and occluding-object descriptions in spoof prompts (ablation) | SLIP using $\mathcal{L}_{L}+\mathcal{L}_{S}$ · OCM→I | HTER (%): 30.92 -> 26.64 | Helped |
| use SLIP instead of OC-SCMNet | one-class FAS · OULU-NPU protocols 1–4 | ACER (%): 23.49 -> 14.58; 25.43 -> 22.67; 23.83$\pm$3.14 -> 23.19$\pm$2.86; 14.04$\pm$4.9 -> 12.96$\pm$5.72 | Helped |
| use SLIP instead of OC-SCMNet | one-class FAS · cross-domain seen attack protocols | HTER (%): 24.05 -> 18.81; 28.02 -> 23.89; 21.36 -> 15.71; 34.37 -> 29.15 | Helped |
| use SLIP instead of OC-SCMNet | one-class FAS · unseen physical attacks | HTER (%): 28.99 -> 19.77; 14.33 -> 7.02; 8.40 -> 3.86 | Helped |
Limitations
- Relatively large model size (171.72M params) due to CLIP text and image encoders during training.
- Detecting small physical adversarial attacks (funny-eye glasses) remains highly challenging.
Experimental setup
Datasets: OULU-NPU CASIA-MFSD MSU-MFSD Idiap Replay-Attack 3DMAD HKBU-MARs CASIA-SURF PADISI-Face
Benchmarks: OULU-NPU OCI-M/OMI-C/OCM-I/ICM-O cross-domain protocols unseen 3D mask/print/replay protocols PADISI-Face physical adversarial protocols
Metrics: APCER BPCER ACER HTER AUC #param FPS
Baselines: IQM-GMM OC-fPAD OC-LCFAS AAE OC-SCMNet IADG SAFAS
Builds on: CLIP OC-SCMNet one-class FAS with generative feature learning
Closest papers by meaning
- One-Class Face Anti-spoofing via Spoof Cue Map-Guided Feature Learning (CVPR 2024 · similarity 0.94)
- FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.91)
- Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models (AAAI 2025 · similarity 0.90)
- From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.89)
- mmFAS: Multimodal Face Anti-Spoofing Using Multi-Level Alignment and Switch-Attention Fusion (AAAI 2025 · similarity 0.88)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · 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.