Research map / Continual learning / Domain and prompt-based
Domain-Hallucinated Updating for Multi-Domain 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 Domain-Hallucinated Updating (DHU), which stores domain-style buffers from previous domains and hallucinates pseudo-previous features for new-domain face anti-spoofing updates without access to old data.
Problem
Multi-domain face anti-spoofing: updating an FAS model on new domains using only new data while avoiding catastrophic forgetting of previous domains (plasticity-stability dilemma), given privacy limits on old data.
Approach
Learns a compact buffer of domain-style statistics (per-layer feature mean/variance) for previous domains, then hallucinates pseudo-previous features from new-domain data by restyling shallow features with a selected stored style, and jointly aligns new and pseudo-previous real-sample features via contrastive and JS-divergence losses.
Key results
- On FASMD protocols, DHU reaches 87.2% average TPR@FPR=0.5% across A→B–E, outperforming the best prior method by about 0.3 to 3.1 points per protocol as reported.
- On the CASIA→OULU→Idiap→MSU long sequence, DHU achieves 84.6 AUC and 20.8 ACER mean, with forgetting ACER 9.6, the lowest among compared methods.
- In the generalization setting (train A,B; test C-E), DHU averages 83.0 TPR and 5.3 ACER, the best TPR and tied-best ACER among the methods listed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hallucinated-feature loss Lhal (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 78.9 -> 82.4 | Helped |
| Add sample-based alignment loss Lsmp (ablation) | DHU with Lcls and Lhal · FASMD A→C | TPR@FPR=0.5% avg: 82.4 -> 84.3 | Helped |
| Add distribution-based alignment loss Ldist (ablation) | DHU with Lcls and Lhal · FASMD A→C | TPR@FPR=0.5% avg: 82.4 -> 84.4 | Helped |
| Add Ldist when Lcls, Lhal, and Lsmp are present (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 84.3 -> 85.5 | Helped |
| Add Lsmp when Lcls, Lhal, and Ldist are present (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 84.4 -> 85.5 | Helped |
| Add Lhal and Lsmp when Lcls and Ldist are present (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 81.4 -> 85.5 | Helped |
| Add Lcls when Lhal, Lsmp, and Ldist are present (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 81.9 -> 85.5 | Helped |
| Use the DHU domain-information learning strategy instead of Random storage (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 71.1 -> 84.3 | Helped |
| Use the DHU domain-information learning strategy instead of K-means (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 78.9 -> 84.3 | Helped |
| Use the DHU domain-information learning strategy instead of DeepCluster (ablation) | DHU · FASMD A→C | TPR@FPR=0.5% avg: 81.9 -> 84.3 | Helped |
| Use DHU instead of ER | ResNet-18 · FASMD unseen domains C-E | TPR@FPR=0.5% avg: 82.4 -> 83.0 | Helped |
| Use DHU instead of Joint training | ResNet-18 · CASIA→OULU→Idiap→MSU | AUC mean: 86.8 -> 84.6 | Hurt |
Surprising findings
- Response regularization methods like LwF and MAS perform worse than simple finetuning on some protocols, attributed to untrustworthy responses on new domains.
Limitations
- Relies on FASMD protocol and a small set of benchmark domains; generality to other domain shifts is not shown.
- Assumes a fixed buffer size of 200 and K-means merging, with the choice of statistics only in shallow layers.
Experimental setup
Datasets: SiW SiW-Mv2 OULU-NPU FASMD CASIA-FASD Idiap Replay-Attack MSU-MFSD
Benchmarks: FASMD (A→B, A→C, A→D, A→E) CASIA→OULU→Idiap→MSU long sequence
Metrics: TPR@FPR=0.5% AUC ACER APCER BPCER
Baselines: Finetune LwF LwM MAS FAS-wrapper ER DER GEM A-GEM Serial Res-Adapter Parallel Res-Adapter
Builds on: Gumbel-Softmax MAS FAS-wrapper Domain information as mean and variance statistics (Wang et al. 2022b)
Closest papers by meaning
- Gradient Alignment for Cross-Domain Face Anti-Spoofing (CVPR 2024 · similarity 0.89)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.88)
- SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.86)
- FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.86)
- mmFAS: Multimodal Face Anti-Spoofing Using Multi-Level Alignment and Switch-Attention Fusion (AAAI 2025 · similarity 0.86)
- Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space (CVPR 2024 · similarity 0.86)
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.