PaperlayerView exact revision on arXiv

Source-linked AI summary

Generalizable Representation Learning for Mixture Domain Face Anti-Spoofing

Zhihong Chen, Taiping Yao, Kekai Sheng, Shouhong Ding, Ying Tai, Jilin Li, Feiyue Huang, Xinyu Jin

arXiv:2105.02453v1cs.CV

TL;DR

Face anti-spoofing DG methods rely on domain labels, yet practical mixture-domain data may not provide them and dataset-based labels may be suboptimal. D2AM iteratively clusters mixture domains using discriminative representations and trains with meta-learning. The paper reports that it outperforms conventional DG-based face anti-spoofing methods, including methods using domain labels.

  • Problem

    DG-based face anti-spoofing methods assume known domain labels, although mixture-domain data may have unknown labels and dataset-based labels may be a suboptimal partition.

    Source: Problem Definition and Notations · p. 2

  • Method

    D2AM iteratively clusters mixture domains using discriminative domain representations and trains a generalizable face anti-spoofing model with meta-learning.

    Source: Conclusion · pp. 7–8

  • Results

    D2AM outperforms conventional DG-based face anti-spoofing methods, including those utilizing domain labels.

    Source: Result and Discussion · p. 6, Conclusion · p. 8

  • Takeaways & Limitations

    Visualizations are presented to enhance interpretability of the proposed method.

    Source: Conclusion · p. 8

    The method assumes that an image's domain information can be represented by its style.

    Source: Domain Feature Define · p. 3

Abstract

from arXiv · show

Face anti-spoofing approach based on domain generalization(DG) has drawn growing attention due to its robustness forunseen scenarios. Existing DG methods assume that the do-main label is known.However, in real-world applications, thecollected dataset always contains mixture domains, where thedomain label is unknown. In this case, most of existing meth-ods may not work. Further, even if we can obtain the domainlabel as existing methods, we think this is just a sub-optimalpartition. To overcome the limitation, we propose domain dy-namic adjustment meta-learning (D2AM) without using do-main labels, which iteratively divides mixture domains viadiscriminative domain representation and trains a generaliz-able face anti-spoofing with meta-learning. Specifically, wedesign a domain feature based on Instance Normalization(IN) and propose a domain representation learning module(DRLM) to extract discriminative domain features for cluster-ing. Moreover, to reduce the side effect of outliers on cluster-ing performance, we additionally utilize maximum mean dis-crepancy (MMD) to align the distribution of sample featuresto a prior distribution, which improves the reliability of clus tering. Extensive experiments show that the proposed methodoutperforms conventional DG-based face anti-spoofing meth-ods, including those utilizing domain labels. Furthermore, weenhance the interpretability through visualizatio

Introduction

Face anti-spoofing methods that perform well within a dataset can degrade in cross-domain settings. Existing DG approaches rely on domain labels, but mixture-domain data may lack them, and dataset-based labels may be a suboptimal partition.

Introduction

  • Existing DG-based face anti-spoofing methods use source-domain labels, which may be unavailable in practical mixture-domain datasets.Manually assigning labels can be expensive and time-consuming; factors such as illumination, background, and camera type also complicate domain definition.Source: Introduction · p. 1
  • Even available dataset-based domain labels may create a suboptimal partition when samples across datasets overlap in distribution.Small distribution differences between domains may keep the model from focusing on better-generalized learning directions.Source: Introduction · pp. 1–2
  • D2AM iteratively divides mixture domains using discriminative domain representations, then trains a generalizable face anti-spoofing model with meta-learning.Its DRLM extracts domain-discriminative features, while MMD-based regularization reduces the distance between sample-feature and prior distributions to limit outlier effects on clustering.Source: Introduction · p. 2
  • The paper frames mixture-domain face anti-spoofing as a scenario and reports experiments and visualizations to evaluate D2AM against state-of-the-art competitors.The introduction identifies DRLM and MMD-based regularization as components for dynamic adjustment.Source: Introduction · p. 2

Related Work

Prior face anti-spoofing work spans handcrafted and deep-learning methods, while DG approaches address performance degradation in unseen scenarios. Related DG research includes domain alignment, meta-learning, and clustering, with D2AM emphasizing task-information distillation for domain clustering.

Face Anti-Spoofing

  • Handcrafted texture cues and deep-learning approaches are used to classify real and spoof faces, but performance degrades in unseen scenarios.Examples include LBP, HOG, SURF, SIFT, CNN features, and CNN-RNN architectures using face-depth and rPPG auxiliary supervision.Source: Face Anti-Spoofing · p. 2
  • DG face anti-spoofing methods learn domain-invariant representations or use adversarial learning and meta-learning, but the cited meta-learning approach requires domain labels.Shao et al. use multi-adversarial DG; Jia et al. use single-side adversarial learning and asymmetric triplet loss; Shao, Lan, and Yuen (2020) explore meta-learning with domain knowledge.Source: Face Anti-Spoofing · p. 2

Deep Domain Generalization

  • Deep DG work aligns feature distributions across source domains or uses meta-learning, while Matsuura and Harada (2020) cluster latent domains but may retain task-discriminative information in domain features.The paper positions its DRLM with style enhancement entropy loss as encouraging domain-discriminative features for domain division.Source: Deep Domain Generalization · p. 2

Our Approach

D2AM targets mixture-domain face anti-spoofing, where domain labels are unavailable, by alternating pseudo-domain assignment and meta-learning. It builds domain-discriminative representations for clustering and uses MMD regularization during training.

Our Approach

Dynamically Assigning Pseudo Domain Labels

  • Domain features concatenate per-channel means and standard deviations from multiple convolutional layers, using feature statistics that IN normalizes to alleviate domain discrepancy.Positive and negative samples are clustered separately, then matched into domains; ImageNet-pretrained ResNet features initialize clustering, and Kuhn-Munkres matching limits label shifts between epochs.Source: Domain Feature Define · p. 3, Clustering with Discriminative Domain… · p. 4
  • DRLM uses channel attention to separate task-discriminative F + features from domain-discriminative F − features, then applies reverse-entropy regularization to suppress task information in F −.The loss constrains the positive-sample probability toward 0.5 for both positive and negative samples, supporting domain clustering.Source: Domain Representation Learning Module · pp. 3–4

MMD-Based Regularized Meta-Learning

  • MMD regularization narrows the distance between sample-feature and prior distributions to reduce outlier effects on clustering and encourage correction of unseen-sample distributions.Training randomly assigns K−1 pseudo-domains as meta-train domains and the remaining domain as meta-test; inner updates include classification and MMD losses, with depth maps also used as auxiliary guidance.Source: MMD-Based Regularized Meta-Learning · p. 4, §17 · p. 4, Meta-Test · p. 5

Experiments

Across the evaluated tasks, D2AM is reported to outperform state-of-the-art methods on most tasks. Ablations and visualizations examine its domain clustering, feature choices, and changing pseudo-domains.

Experiments

  • Three of OULU-NPU, CASIA-FASD, Idiap Replay-Attack, and MSU-MFSD form unlabeled mixture sources; the fourth is the unseen test domain.HTER and AUC are the evaluation metrics.Source: Datasets · p. 5, Implementation Details · p. 5

Result and Discussion

  • D2AM outperforms state-of-the-art methods on most tasks, including domain-generalization methods that use domain labels.The authors also report stronger average results than RFM on Oulu-NPU protocol 3, but provide no overall cross-task values here.Source: Result and Discussion · p. 6, Table 2 · p. 6
  • On O&C&M to I, domain-labeled splits outperform random splits, while low-attention-only clustering outperforms using all channels.The authors also report the best results when K is set based on clusters from pre-trained ResNet features; D2AM performs best among the tested components.Source: Ablation Study · p. 6, Table 3 · p. 6

Visualization and Analysis

  • In adaptation-feature t-SNE visualizations, D2AM shows greater domain invariance, more dispersed task features than RFM, and no outliers under MMD regularization.Figure 4 visualizes the O&C&M to I task; its domain labels follow RFM rather than D2AM's pseudo-labels.Source: Adaptation Feature Visualization · p. 6, Figure 4 · p. 6
  • DRLM visualizations distinguish feature roles: F+ contains more task-discriminative information, while F− contains more domain information.The authors say domain enhancement entropy loss Lp encourages extraction of domain-discriminative features.Source: Convolutional Feature Visualization · pp. 6–7
  • Low-attention maps focus on attack cues, backgrounds, and styles, while high-attention maps focus on the internal face; low-weight features are used for domain division.The authors note that low-attention cues may change in a new scene, whereas internal-face features are more likely to be intrinsic and generalized.Source: Attention Map Visualization · p. 7, Figure 5 · p. 7
  • Cluster visualizations show changing division bases: Epoch 3 clusters mainly by illumination, while Epoch 6 shifts mainly to background.The authors attribute the change to meta-learning improving robustness to the current shift, so subsequent clustering targets other scene information not yet robustly learned.Source: Cluster Visualization · p. 7, Figure 6 · p. 7
  • Pseudo-domain NMI with previous assignments is about 0.6–0.8, and pseudo-domain labels overlap actual domains by only 64.9%±5.7%; MMD shows larger differences between pseudo-subdomains than actual domains.Figure 7 reports the MMD and NMI comparisons on O&C&M to I.Source: Cluster Analysis · p. 7, Figure 7 · p. 7

Conclusion

D2AM addresses mixture-domain face anti-spoofing without domain labels by iteratively clustering domains through discriminative representations and training with meta-learning. The authors report that it outperforms conventional DG-based methods, including methods using domain labels, and provide visualizations to enhance interpretability.

Conclusion

  • D2AM is presented as the first work to address mixture-domain face anti-spoofing when domain labels are unknown.It iteratively clusters mixture domains via discriminative domain representations and trains a generalizable feature extractor by meta-learning.Source: Conclusion · p. 7
  • D2AM outperforms conventional DG-based face anti-spoofing methods, including those that use domain labels.The authors also report visualizations to enhance interpretability.Source: Conclusion · p. 8
  • D2AM uses a DRLM and MMD-based regularization for dynamic adjustment to simulate more difficult and abundant domain-shift scenes.Source: Conclusion · p. 8
Loading 2105.02453v1…