Research map / Label-efficient and robust learning / Recognition and long-tail
Mixture-of-Attack-Experts with Class Regularization for Unified Physical-Digital Face Attack Detection
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 MoAE-CR, a CLIP-based unified physical-digital face attack detector that adds multi-head Mixture-of-Attack-Experts (Soft MoE variant) to the image encoder plus Disentanglement and Cluster Distillation losses for class-aware regularization.
Problem
Unified detection of digital (deepfake/editing/adversarial) and physical (print/replay) face attacks, where intra-class variation across attack types is large and inter-class variation between live and fake faces is small.
Approach
Combines a multi-head Soft MoE image encoder in CLIP (MoAE) with two batch-wide class-center losses (DM pushes live/fake centers apart; CDM pulls features to their class center and away from other centers), both using a log-sum-exp weighting that emphasizes distant/hard features.
Key results
- On UniAttackData Protocol 1, MoAE-CR reaches ACER 0.37%, ACC 99.47%, AUC 99.97%, EER 0.49%, beating UniAttackDetection (ACER 0.52%).
- On UniAttackData Protocol 2, average ACER 15.13% vs 22.42% for the best prior method (UniAttackDetection), with ACC 85.41%.
- On JFSFDB (P2 intra), ACER 4.40% vs 7.70% for ResNet50 (next best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SoftMoE to CLIP (ablation) | CLIP · UniAttackData Protocol I | ACER (%): 0.79 -> 0.66 | Helped |
| Replace SoftMoE with multi-head MoAE (ablation) | CLIP · UniAttackData Protocol I | ACER (%): 0.66 -> 0.49 | Helped |
| Add DM without CDM to CLIP + MoAE (ablation) | CLIP + MoAE · UniAttackData Protocol I | ACER (%): 0.49 -> 0.95 | Hurt |
| Add CDM without DM to CLIP + MoAE (ablation) | CLIP + MoAE · UniAttackData Protocol I | ACER (%): 0.49 -> 0.54 | Hurt |
| Add DM and CDM together to CLIP + MoAE (ablation) | CLIP + MoAE · UniAttackData Protocol I | ACER (%): 0.49 -> 0.37 | Helped |
| Add DM and CDM together to CLIP + MoAE (ablation) | CLIP + MoAE · UniAttackData Protocol I | ACC (%): 99.24 -> 99.47 | Helped |
| Add DM and CDM together to CLIP + MoAE (ablation) | CLIP + MoAE · UniAttackData Protocol I | AUC (%): 99.79 -> 99.97 | Helped |
| Increase attention heads or experts beyond the selected configuration (ablation) | MoAE · UniAttackData Protocol I | ACER (%): 0.37; 0.41; 0.52; 0.67; 1.09; 1.60 | Mixed |
| Use MoAE-CR instead of UniAttackDetection | face attack detection · UniAttackData Protocol 1 | ACER (%): 0.52 -> 0.37 | Helped |
| Use MoAE-CR instead of UniAttackDetection | face attack detection · UniAttackData Protocol 2 | ACC (%): 67.35±23.22 -> 85.41±6.85 | Helped |
| Use MoAE-CR instead of ResNet50 | face attack detection · JFSFDB P2 intra | ACER (%): 7.70 -> 4.40 | Helped |
Surprising findings
- Adding DM alone to the MoAE model hurt performance (ACER rose from 0.49% to 0.95%), and CDM alone gave only modest gains, with gains appearing only when both are combined.
Limitations
- Improvements in Protocol 2 are small relative to large standard deviations (e.g., 15.13±12.10 ACER).
- Hyperparameters (threshold t=0.5, experts/heads) selected on the test-reported benchmark with limited sensitivity analysis.
- Hardware and training compute are not reported.
Experimental setup
Datasets: UniAttackData JFSFDB
Benchmarks: UniAttackData Protocol 1 UniAttackData Protocol 2 (P2.1/P2.2) JFSFDB
Metrics: ACER ACC AUC EER
Baselines: ResNet50 ViT-B/16 Auxiliary (Depth) CDCN FFD UniAttackDetection SupContrastive N-pair Triplet Hard Triplet
Builds on: CLIP Soft MoE triplet loss N-pair loss supervised contrastive loss
Closest papers by meaning
- One-Class Face Anti-spoofing via Spoof Cue Map-Guided Feature Learning (CVPR 2024 · similarity 0.85)
- FSFM: A Generalizable Face Security Foundation Model via Self-Supervised Facial Representation Learning (CVPR 2025 · similarity 0.84)
- SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.83)
- FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.83)
- Multi-Domain Incremental Learning for Face Presentation Attack Detection (AAAI 2024 · similarity 0.83)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.82)
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.