Research map / Label-efficient and robust learning / Recognition and long-tail

One-Class Face Anti-spoofing via Spoof Cue Map-Guided Feature Learning

CVPR 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Fair and quality-aware face recognition · 21 citations (OpenAlex)

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 OC-SCMNet, a one-class face anti-spoofing model that learns spoof cue maps (zero for live faces) guided by pseudo SCMs, a conditional generator of latent spoof features, and a memory bank encouraging exploration of latent space.

Problem

Face anti-spoofing when only live face images are available for training, so the model must detect unseen out-of-distribution spoof attacks.

Approach

Estimate spoof cue maps that should be zero for live images, and use a conditional generator driven by combinations of nonzero pseudo SCMs to synthesize latent spoof features; a memory bank of mutually independent latent features pushes generation into unexplored regions of latent space.

Key results

• On cross-domain C→I, OC-SCMNet reaches 7.29% HTER and 96.68% AUC, versus 30.79% HTER for the baseline without generator and memory loss terms.
• On intra-domain OULU-NPU protocol 1, OC-SCMNet achieves 23.49% ACER, compared with 30.24% for Baweja et al. and 46.95% for IQM-GMM among one-class methods.
• With one-shot spoof supervision, OC-SCMNet reaches 2.14% HTER on C→I and 1.11% on I→C, comparable to two-class methods.

Measured effects reported in the paper

Surprising findings

• Memory bank size beyond 16 yields no further improvement, suggesting overcompleteness.
• Selection thresholds above 0.5 degrade performance sharply, showing that the diversity constraint on generated features matters.

Limitations

• Performance on intra-domain testing remains well below two-class methods.
• One-shot results require access to one spoof image during training, which departs from the strict one-class setting.

Experimental setup

Datasets: OULU-NPU CASIA-MFSD MSU-MFSD Idiap Replay-Attack SiW 3DMAD HKBU-MARs CASIA-SURF

Benchmarks: OULU-NPU CASIA-MFSD MSU-MFSD Idiap Replay-Attack SiW 3DMAD HKBU-MARs CASIA-SURF

Metrics: APCER BPCER ACER HTER AUC

Baselines: IQM-GMM Baweja et al. Lim et al. AAE CDCN PatchNet LDCN IADG SAFAS SSAN-M

Builds on: Spoof cue map estimation (Yu et al. CVPR 2020) Gaussian noise one-class augmentation (Baweja et al. IJCB 2020) AAE (CCBR 2021)

Closest papers by meaning

• SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.94)
• From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.87)
• FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.87)
• Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.86)
• Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models (AAAI 2025 · similarity 0.85)
• mmFAS: Multimodal Face Anti-Spoofing Using Multi-Level Alignment and Switch-Attention Fusion (AAAI 2025 · similarity 0.85)