Research map / Continual learning / Domain and prompt-based
Multi-Domain Incremental Learning for Face Presentation 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 MDIL-PAD, a buffer-free multi-domain incremental learning method for face presentation attack detection using adaptive domain-specific ViT experts, an instance-wise router, and an asymmetric classifier.
Problem
Face presentation attack detection must learn new attack domains sequentially without access to previous domain data, where fine-tuning causes catastrophic forgetting and domain index is unknown at test time.
Approach
Adaptive Domain-specific Experts in the last ViT blocks (per-domain MLP experts with shared attention), an Instance-wise Router that assigns test instances to a domain expert via similarity to learned domain centers, and a single asymmetric classifier with a shared live class and per-domain spoof classes to keep predicted probability distributions consistent.
Key results
- In 4-domain incremental setting (M→C→I→O), the method reaches avg Δm% of 1.17, versus 16.94 for fine-tuning and 2.18 for S-iPr, with M-domain HTER 6.52 (vs FT 21.92 at step 2 relative to JT 1.71).
- On cross-dataset LOO benchmarks, the IL method reports HTER 5.71% on O&C&I to M, close to DG/DA methods (e.g., SSAN-R 6.67%, GDA 9.20%).
- Cross-attack-type SiW/OULU setting reduces degradation by 16.33% (S-P→O-R) and 11.64% (O-P→S-R) relative to FT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ADE experts with the instance-wise router and asymmetric classifier instead of fine-tuning | MDIL-PAD · SiW-Print→OULU-Replay, step 2 | HTER (%): 31.17 -> 0.65 | Helped |
| Use ADE experts with the instance-wise router and asymmetric classifier instead of fine-tuning | MDIL-PAD · OULU-Print→SiW-Replay, step 2 | HTER (%): 28.20 -> 6.65 | Helped |
| Use ADE experts with the instance-wise router and asymmetric classifier instead of fine-tuning | MDIL-PAD · M→C→I→O, step 4; prior domains M, C, I | HTER (%): 33.94 -> 10.62; 25.56 -> 2.54; 0.64 -> 0.31 | Mixed |
| Use MDIL-PAD for cross-domain testing rather than SSAN-R | MDIL-PAD · O&C&I→M | HTER (%): 6.67 -> 5.71 | Helped |
| Use MDIL-PAD for cross-domain testing rather than SSAN-R | MDIL-PAD · O&M&I→C | HTER (%): 10.00 -> 13.22 | Hurt |
| Use MDIL-PAD for cross-domain testing rather than SSAN-R | MDIL-PAD · O&C&M→I | HTER (%): 8.88 -> 11.25 | Hurt |
| Use MDIL-PAD for cross-domain testing rather than SSAN-R | MDIL-PAD · I&C&M→O | HTER (%): 13.72 -> 12.47 | Helped |
| Replace separate expert classifiers with an asymmetric classifier (ablation) | MDIL-PAD · M, step 4, domain-agnostic | HTER (%): 24.70 -> 10.62 | Helped |
| Replace separate expert classifiers with an asymmetric classifier (ablation) | MDIL-PAD · M, step 4, domain-aware | HTER (%): 9.68 -> 6.33 | Helped |
| Move the IwR insertion position to the 8th block (ablation) | MDIL-PAD · M→C→I→O, step 4 | HTER (%) | Helped |
Surprising findings
- Under the stringent sequential-training setting, the IL method approximates or outperforms some DA/DG methods on cross-dataset LOO benchmarks.
- Multi-classifier designs with per-expert thresholds produce inconsistent predicted probability distributions, causing performance drops even when each expert is accurate.
Limitations
- Domain index must be known during training and the router relies on domain-center similarity at test time, which may be fragile for unseen domains.
- Sequential training theoretically limits generalizable representation learning; the method still trails joint-training upper bounds (e.g., JT avg Δm% reference).
Experimental setup
Datasets: OULU-NPU CASIA-MFSD Idiap ReplayAttack MSU-MFSD SiW
Benchmarks: OULU-NPU CASIA-MFSD Idiap ReplayAttack MSU-MFSD SiW
Metrics: HTER AUC Delta m% (average relative HTER drop)
Baselines: Fine-tuning (FT) Feature extraction (FE) LwF DyTox L2P S-iPr SDA VLAD GDA SSDG-R SSAN-R
Builds on: ViT LwF DyTox L2P S-Prompts
Closest papers by meaning
- Non-exemplar Domain Incremental Object Detection via Learning Domain Bias (AAAI 2024 · similarity 0.84)
- Domain-Hallucinated Updating for Multi-Domain Face Anti-spoofing (AAAI 2024 · similarity 0.84)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.84)
- Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework (AAAI 2024 · similarity 0.83)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.83)
- Link-based Contrastive Learning for One-Shot Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.83)
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.