Source-linked AI summary
Regularized Fine-grained Meta Face Anti-spoofing
Rui Shao, Xiangyuan Lan, Pong C. Yuen
TL;DR
Face anti-spoofing methods often generalize poorly to unseen attacks because they learn dataset-biased cues. This paper proposes regularized fine-grained meta-learning, combining domain-knowledge-regularized features with multiple simulated domain shifts. Experiments on four public datasets validate the method's effectiveness.
Problem
Face anti-spoofing methods often learn dataset-biased differentiation cues and therefore generalize poorly to unseen attacks.
Method
The framework performs meta-learning in a feature space regularized by face depth supervision and simultaneously trains across varied simulated domain-shift scenarios.
Results
Experiments on four public datasets validate the effectiveness of the proposed regularized fine-grained meta-learning framework.
Takeaways & Limitations
The framework is designed to improve generalization to unseen attacks by learning more generalized differentiation cues.
Abstract
from arXiv · showhide
Face presentation attacks have become an increasingly critical concern when face recognition is widely applied. Many face anti-spoofing methods have been proposed, but most of them ignore the generalization ability to unseen attacks. To overcome the limitation, this work casts face anti-spoofing as a domain generalization (DG) problem, and attempts to address this problem by developing a new meta-learning framework called Regularized Fine-grained Meta-learning. To let our face anti-spoofing model generalize well to unseen attacks, the proposed framework trains our model to perform well in the simulated domain shift scenarios, which is achieved by finding generalized learning directions in the meta-learning process. Specifically, the proposed framework incorporates the domain knowledge of face anti-spoofing as the regularization so that meta-learning is conducted in the feature space regularized by the supervision of domain knowledge. This enables our model more likely to find generalized learning directions with the regularized meta-learning for face anti-spoofing task. Besides, to further enhance the generalization ability of our model, the proposed framework adopts a fine-grained learning strategy that simultaneously conducts meta-learning in a variety of domain shift scenarios in each iteration. Extensive experiments on four public datasets validate the effectiveness of the proposed method.
Introduction
Face anti-spoofing must generalize beyond dataset-specific attacks, but existing methods often learn biased cues. The paper addresses this domain-generalization challenge with regularized fine-grained meta-learning.
- Motivation: Face recognition systems face increasingly critical risks from print and video replay attacks that can resemble genuine faces.These attacks exploit printed photos or digital images and videos on devices such as mobile phones and laptops.
- Limitations of Existing Methods: Existing appearance- and temporal-based methods perform promisingly within datasets but degrade when testing data come from different distributions.Cross-domain shifts mainly reflect different attack materials or recording environments.
- Problem Formulation: The paper formulates face anti-spoofing as domain generalization, which assumes no access to target-domain information.This differs from unsupervised domain adaptation, which uses unlabeled target-domain data.
- Problem Formulation: Vanilla meta-learning can fail because binary supervision encourages arbitrary differentiation cues and typically simulates only one domain-shift scenario per iteration.The paper identifies both limitations as obstacles to generalization across unseen attacks.
- Proposed Direction: Regularized fine-grained meta-learning uses anti-spoofing domain knowledge to guide feature learning while jointly training across multiple domain-shift scenarios.The framework seeks coordinated, better-generalized learning directions from meta-train and meta-test domains.
Related Work
Prior face anti-spoofing methods rely on appearance or temporal cues but often fail under cross-dataset attacks. The proposed method instead targets generalized cues through simulated domain shifts without adversarial discriminators.
- Face Anti-spoofing Methods: Face anti-spoofing research commonly uses appearance-based or temporal-based cues to distinguish real and fake faces.Appearance methods include texture and distortion descriptors, while temporal methods exploit changes across frames.
- Generalization Limitation: Both method families degrade in cross-dataset tests because their learned cues can be biased toward attack materials or recording environments.The resulting limitation appears when unseen attacks are encountered.
- Domain Generalization: Compared with adversarial domain-generalization methods, this approach uses meta-learning under simulated domain shifts without assuming a shared discriminative feature space or training extra discriminators.The paper presents this design as more efficient.
Proposed Method
The framework simulates domain shifts by splitting source domains into meta-train and meta-test sets, while depth supervision regularizes feature learning. Fine-grained updates jointly exploit multiple cross-domain scenarios.
- Domain Shift Simulation: At each iteration, N−1 source domains are randomly selected for meta-training and the remaining domain is used for meta-testing.This constructs simulated domain shifts from the available source domains.
- Domain Knowledge Regularization: The method uses face depth maps as domain knowledge because live faces have facial depth whereas flat attacks are treated as having none.PRNet estimates depth maps for real faces, and zero maps supervise fake faces.
- Meta-Train and Meta-Test: For each meta-train domain, the meta learner is updated using binary-label cross-entropy before evaluation on the disjoint meta-test domain.The updated meta learners are jointly evaluated through multiple cross-entropy classifications.
- Meta-Optimization: Meta-optimization jointly trains the feature extractor, depth estimator, and meta learner using information from meta-train and meta-test losses.The objective also coordinates meta-train and meta-test gradient directions to reduce overfitting to a single domain.
- Domain Knowledge Regularization: Depth-estimation regression losses provide auxiliary supervision to the feature extractor, so classification is performed in a domain-knowledge-regularized feature space.This regularization is intended to focus meta-training and meta-testing on better-generalized learning directions.
Experiments
Experiments across four public datasets evaluate domain generalization to an unseen dataset using HTER and AUC, comparing the proposed method with conventional and meta-learning baselines. Results and ablations support the contributions of regularization, meta-learning, fine-grained scenarios, and second-order information.
- Experimental Setup: Experiments use Oulu-NPU, CASIA-MFSD, Idiap Replay-Attack, and MSU-MFSD, treating each dataset as a domain with three sources and one unseen test domain.HTER and AUC are the evaluation metrics.
- Comparison Results: The proposed method outperforms state-of-the-art face anti-spoofing methods on four domain-generalization testing sets.The comparison includes appearance-, temporal-, and auxiliary-supervision-based baselines.
- Comparison Results: The proposed method also outperforms vanilla meta-learning for domain generalization methods on the face anti-spoofing task.The authors attribute this to addressing the two identified issues through regularized and fine-grained meta-learning.
- Components Evaluation: Removing either meta-learning or domain-knowledge regularization degrades performance in the O&M&I-to-C evaluation.The ablation evaluates the proposed network without meta-learning and without face-depth-map regularization.
- Fine-Grained Learning and Derivatives: The method evaluates fine-grained learning and second-order derivative information through additional experiments, with the second-order variant performing better than the first-order approximation.The first-order comparison replaces the second-order computation with the approximation proposed in Reptile.
- Attention Map Visualization: Attention visualizations show Binary CNN focusing on backgrounds or paper edges and fingers, whereas the proposed method focuses on the internal face region for unseen attacks.The paper presents these regions as more or less generalized differentiation cues, respectively.
Conclusion
The paper frames face anti-spoofing as domain generalization and proposes regularized fine-grained meta-learning to improve generalization to unseen attacks. Its framework uses domain-knowledge supervision and multiple simulated domain-shift scenarios, with experiments validating effectiveness.
- The proposed framework casts face anti-spoofing as a domain generalization problem.
- Meta-learning occurs in a feature space regularized by face anti-spoofing domain-knowledge supervision.
- Fine-grained learning simultaneously exploits varied domain-shift scenarios to train generalization to unseen attacks.
- Comprehensive experiments validate the proposed method statistically and visually.
Supplementary Material
The supplementary material covers four face anti-spoofing datasets, cross-dataset variations, and the proposed network’s structural details. The datasets include both print and video replay attacks, while the network processes RGB and HSV inputs.
- Table 4 compares the four experimental datasets.
- The evaluation uses Oulu-NPU, CASIA-MFSD, Idiap Replay-Attack, and MSU-MFSD datasets containing print and video replay attacks.
- The network uses 256 × 256 × 6 inputs formed from RGB and HSV channels.
- Cross-dataset variations arise from differences in materials, illumination, background, resolution, and related factors.
- Table 5 details the structure of the feature extractor, meta learner, and depth estimator components.