Source-linked AI summary
Self-Domain Adaptation for Face Anti-Spoofing
Jingjing Wang, Jingyi Zhang, Ying Bian, Youyi Cai, Chunmao Wang, Shiliang Pu
TL;DR
Face anti-spoofing methods can generalize poorly to unseen domains, while existing adaptation approaches either require target data during training or do not use target-domain information. This paper meta-learns an adaptor from multiple source domains and updates it using unlabeled test data at inference. Experiments report promising results across four public datasets and performance above most state-of-the-art methods.
Problem
Face anti-spoofing methods generalize poorly to unseen attacks, and existing DA or DG approaches respectively face unknown target domains during training or leave target-domain information unused.
Source: Related Work · p. 2, Conclusion · p. 7Method
The method meta-learns an adaptor from multiple source domains, then updates it with an unsupervised loss using only unlabeled test-domain data at inference.
Source: Learning Adaptor at Training · pp. 3–4Results
The method reports promising results on four public datasets, outperforming most state-of-the-art methods and all compared self-domain adaptation methods, with slightly worse results than RFM.
Source: Experimental Comparison · pp. 5–6, Conclusion · p. 7Takeaways & Limitations
The authors propose using domain-specific test information to extract discriminative features and further boost face anti-spoofing performance.
Source: Conclusion · p. 7
Abstract
from arXiv · showhide
Although current face anti-spoofing methods achieve promising results under intra-dataset testing, they suffer from poor generalization to unseen attacks. Most existing works adopt domain adaptation (DA) or domain generalization (DG) techniques to address this problem. However, the target domain is often unknown during training which limits the utilization of DA methods. DG methods can conquer this by learning domain invariant features without seeing any target data. However, they fail in utilizing the information of target data. In this paper, we propose a self-domain adaptation framework to leverage the unlabeled test domain data at inference. Specifically, a domain adaptor is designed to adapt the model for test domain. In order to learn a better adaptor, a meta-learning based adaptor learning algorithm is proposed using the data of multiple source domains at the training step. At test time, the adaptor is updated using only the test domain data according to the proposed unsupervised adaptor loss to further improve the performance. Extensive experiments on four public datasets validate the effectiveness of the proposed method.
Related Work
Earlier face anti-spoofing work spans texture and temporal cues, while domain adaptation and generalization address cross-domain performance. Self-domain adaptation is related but uses unlabeled test-domain data during inference without requiring target data during training.
Traditional Face Anti-spoofing Methods
- Texture-based methods use handcrafted or learned cues, while temporal-based methods use motions or rPPG signals to detect spoofing.Examples include LBP, HoG, SIFT, SURF, mouth motion, and eye blinking.Source: Related Work · p. 2
- Despite strong intra-dataset results, these methods can degrade sharply across datasets because they discard domain relationships and learn dataset-biased features.Source: Related Work · p. 2
DA & DG based Face Anti-spoofing Methods
- DA methods minimize source–target distribution discrepancy, while DG methods learn domain-invariant features without target-domain data.Li et al. and Wang et al. use MMD and adversarial training; Shao et al. use a multi-adversarial DG framework.Source: Related Work · p. 2
- Unlike related approaches, this work meta-learns an adaptor that uses target-domain data at inference, rather than learning domain-invariant features alone.Self-training DA methods assume target data is accessible during training; Shao, Lan, and Yuen (Shao, Lan, and Yuen 2020) use meta-learning for domain-invariant features.Source: Related Work · pp. 2–3
Self-domain adaptation methods
- Prior self-domain adaptation approaches differ in their adaptation requirements, including living-face data or input, feature, and representation transformations.Qin et al. require living faces for adaptation; other approaches generate target-style data, modulate representations, or transform inputs and features.Source: Related Work · p. 3
Proposed Method
The method first learns an adaptor across source domains, then optimizes it with unlabeled test-domain data while keeping the rest of the model fixed. Meta-learning trains the adaptor to adapt to a held-out domain, and the unsupervised loss combines reconstruction, prediction entropy, and an orthogonality constraint.
Overview
- The framework learns an adaptor from multiple source domains, updates it on unlabeled test-domain data, then freezes the model for testing.The source domain is unavailable during inference adaptation.Source: Overview · p. 3, Adapting at Inference · p. 5
Learning Adaptor at Training
- Meta-training simulates domain shift by selecting distinct source domains for labeled meta-training and unlabeled meta-testing batches.The adaptor is integrated through a residual connection to the classifier’s first convolution; removing it restores the original classifier.Source: Learning Adaptor at Training · p. 3
- The target-domain adaptor loss uses source-trained autoencoder reconstruction, prediction entropy, and orthogonality regularization.The reconstruction term measures similarity to source features, entropy encourages confident target predictions, and orthogonality is imposed to prevent feature mode collapse.Source: Learning Adaptor at Training · pp. 3–4
Adapting at Inference
- At inference, only the adaptor is optimized on unlabeled test-domain data; all other parameters remain fixed during adaptation and testing.Source: Adapting at Inference · p. 5
Experiments
Across four cross-dataset testing tasks, the proposed method outperforms most state-of-the-art face anti-spoofing methods and all compared self-domain adaptation methods. Ablations and visualizations examine its loss components, learning steps, and target-domain feature adaptation.
Experimental Settings
- Four testing tasks use three datasets as sources and the remaining dataset as target, with HTER and AUC as evaluation metrics.The datasets are OULU-NPU, CASIA-MFSD, Idiap Replay-Attack, and MSU-MFSD.Source: Experimental Settings · p. 5
Experimental Comparison
- The proposed method outperforms most state-of-the-art methods, with RFM the stated exception, and outperforms all compared self-domain adaptation methods.The comparison covers four testing domains.Source: Experimental Comparison · pp. 5–6, Table 1 · p. 6, Table 2 · p. 6
Ablation Study
- Removing any unsupervised adaptor-loss component decreases performance, and excluding either meta-learning initialization or test-time adaptor optimization degrades results.The relative gains from the two learning steps vary with domain shift: meta-initialization contributes more on O&C&M to I, while further adaptation contributes more on O&M&I to C.Source: Ablation Study · pp. 6–7, Table 3 · p. 7
Visualizations of the Proposed Method
- After adaptation, target-domain C's real and fake face features are more compact and farther apart, while some source-domain features become less separable.The authors note that inference adaptation targets the test domain; the adaptor can be removed for source-domain predictions.Source: Figure 6 · p. 7, Visualizations of the Proposed Method · p. 7
Conclusion
The proposed self-domain adaptation framework uses test-domain information at inference to improve spoofing-face detection. Experiments on four public datasets show promising results.
Conclusion
- Experiments show the method is effective and achieves promising results on four public datasets.Source: Conclusion · p. 7
- A meta-learning-based adaptor learning algorithm provides better adaptor initialization.Source: Conclusion · p. 7
- An adaptor loss enables unsupervised adaptor updates, and the paper proposes extracting discriminative features from test-domain-specific information.Source: Conclusion · p. 7