Source-linked AI summary
Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing
Zhuo Wang, Zezheng Wang, Zitong Yu, Weihong Deng, Jiahong Li, Tingting Gao, Zhongyuan Wang
TL;DR
Generalizable face anti-spoofing is challenged by cross-domain degradation, incomplete use of image statistics, and evaluation protocols that poorly reflect real-world data distributions. The paper proposes SSAN, which separates and reassembles content and style features with adversarial and contrastive learning, alongside a large-scale benchmark. SSAN achieves state-of-the-art performance on existing and proposed benchmarks.
Problem
Cross-domain FAS generalization is limited by dataset bias, underuse of distinct image statistics, and insufficiently realistic evaluation data distributions.
Method
SSAN separates content and style features, uses adversarial learning for domain-indistinguishable content, contrastive learning for liveness-related style, and reassembles correct feature pairs.
Results
SSAN achieves state-of-the-art performance on existing and proposed FAS benchmarks.
Takeaways & Limitations
The work provides a feature-level style-assembly framework and a large-scale benchmark for evaluating generalizable FAS under realistic data distributions.
Takeaways & Limitations
The content-feature approach assumes small distribution discrepancies across domains because facial areas and physical attributes are commonly shared.
Abstract
from arXiv · showhide
With diverse presentation attacks emerging continually, generalizable face anti-spoofing (FAS) has drawn growing attention. Most existing methods implement domain generalization (DG) on the complete representations. However, different image statistics may have unique properties for the FAS tasks. In this work, we separate the complete representation into content and style ones. A novel Shuffled Style Assembly Network (SSAN) is proposed to extract and reassemble different content and style features for a stylized feature space. Then, to obtain a generalized representation, a contrastive learning strategy is developed to emphasize liveness-related style information while suppress the domain-specific one. Finally, the representations of the correct assemblies are used to distinguish between living and spoofing during the inferring. On the other hand, despite the decent performance, there still exists a gap between academia and industry, due to the difference in data quantity and distribution. Thus, a new large-scale benchmark for FAS is built up to further evaluate the performance of algorithms in reality. Both qualitative and quantitative results on existing and proposed benchmarks demonstrate the effectiveness of our methods. The codes will be available at https://github.com/wangzhuo2019/SSAN.
1. Introduction
Face anti-spoofing methods often degrade across domains because limited training data induces dataset bias, while existing DG methods overlook distinct global and local image statistics. SSAN separates content and style, assembles them at feature level, and evaluates generalization with a twelve-dataset benchmark.
- Cross-domain FAS performance can degrade dramatically because limited training data leaves models trapped in dataset bias.
- Existing DG methods operate mainly on complete representations, overlooking the distinct statistical properties of global and local FAS image information.
- SSAN separates content and style features, using adversarial learning for domain-indistinguishable content and contrastive learning to refine liveness-related style.
- Feature-level shuffled style assembly reassembles varied content and style features into a stylized feature space for generalizable FAS.
- A twelve-dataset benchmark models real-world data quantity and distribution, with single-side TPR@FPR proposed for comprehensive assessment.
- The proposed methods achieve state-of-the-art performance on existing and proposed benchmarks.
2. Related Work
Related FAS work spans handcrafted and deep representations, domain adaptation and generalization, normalization and style transfer, and increasingly broad evaluation protocols. SSAN differs by assembling content and style features directly in feature space and evaluating across twelve heterogeneous datasets.
- FAS methods evolved from handcrafted descriptors to deep classifiers augmented with depth, reflection, and rPPG information.
- Domain adaptation and generalization methods learn shared, disentangled, or meta-learned representations to improve performance on unseen domains.
- SSAN splits complete representations into content and style features with different supervision, then reassembles them under contrastive learning.
- Unlike prior normalization and image-level transformation methods, SSAN uses AdaIN-based layers to assemble content and style features in a generalized feature space.
- Most FAS protocols use one or two datasets, limiting evaluation across multiple distributions; broader protocols address this limitation.
- The proposed benchmark fuses twelve image and video datasets with varied domains and attack types, including print, replay, mask, makeup, and waxworks.
3. Proposed Approach
SSAN separates content and style information, recombines them through self-assembly and shuffled style assembly, and uses adversarial and contrastive objectives to improve generalized face anti-spoofing representations.
- Content and Style Information Aggregation: SSAN uses two streams to capture content and style features from different image statistics.Content features represent global semantic and physical information, while style features capture discriminative liveness-related and domain-specific cues.
- Content and Style Information Aggregation: Adversarial learning encourages content features to become indistinguishable across domains.The content generator maximizes the adversarial loss while the domain discriminator optimizes in the opposite direction, with gradient reversal enabling simultaneous optimization.
- Shuffled Style Assembly: AdaIN-based style assembly combines content and style features with residual convolutional processing to form stylized features.The method assembles paired features and uses channel-wise style statistics, affine parameters, and convolution operators within style assembly layers.
- Shuffled Style Assembly: Shuffled style assembly randomly pairs a sample’s content feature with another sample’s style feature to create an auxiliary stylized feature space.Self-assembly uses paired content and style features, whereas shuffle-assembly uses a uniformly chosen permutation of style features.
- Contrastive Learning for Stylized Features: Contrastive learning treats self-assembly features as fixed anchors and guides shuffled features according to liveness-label consistency.A stop-gradient operation fixes the self-assembly anchors, while gradients pass through shuffle-assembly features to emphasize liveness-related style information and suppress domain-specific style information.
4. Large-Scale FAS Benchmarks
The paper introduces a twelve-dataset benchmark to better approximate real-world FAS data quantity and distribution, using protocols that include intra-dataset and challenging cross-domain evaluation.
- Benchmark Motivation: The benchmark merges twelve datasets to address the gap between academic data quantity and authentic scenarios.The authors design corresponding intra- and inter-dataset testing protocols to evaluate FAS methods at larger scale.
- Benchmark Motivation: Real-world evaluation is motivated by live-face prevalence, multiple data fields, and the mismatch between academic testing distributions and deployment conditions.Existing protocols often use nearly balanced live and spoof testing sets and fewer data domains than reality.
- Evaluation Protocols: Protocol 1 evaluates models in an intra-dataset setting using all datasets for training and testing.The passage defines this as the benchmark’s intra-dataset evaluation scenario.
- Evaluation Protocols: Protocol 2 evaluates cross-domain generalization by training on one six-dataset pile and testing on the other, then reversing the direction.The two subprotocols use P1 = {D3, D4, D5, D10, D11, D12} and P2 = {D1, D2, D6, D7, D8, D9}; these tests cover more unseen datasets and complex unknown attacks.
5. Experiments
Experiments evaluate SSAN across cross-domain, limited-source, large-scale, and ablation settings. The results support its generalization, component choices, stable training behavior, and feature-space design.
- Cross-domain evaluation: SSAN-M achieves the best performance on three leave-one-out protocols and competitive performance on the fourth, demonstrating cross-dataset generalization.SSAN-R further improves performance and exceeds SSDG-R under similar settings.
- Limited source domains: SSAN achieves the lowest HTER and highest AUC when trained with only two source domains and tested on the remaining datasets.The result supports modeling efficiency and generalization under limited source data.
- Large-scale benchmark: SSAN achieves the best performance on the proposed large-scale benchmark, while some methods strong on existing protocols sharply degrade there.The benchmark combines twelve datasets to better reflect real-world data quantity and distribution.
- Component analysis: Removing adversarial learning tests whether generalized content features reduce domain differences and support arbitrary content-style assemblies.The experiment targets domain distinction in content information rather than directly operating on complete features.
- Component analysis: Contrastive learning improves cross-domain FAS by guiding style assembly with liveness-intensive cues.The ablation compares SSAN with a version without the contrastive loss.
- Ablation analysis: Stop-gradient improves feature aggregation, emphasizes liveness-related style information, suppresses domain-specific information, and contributes to stable training.The comparison uses continuous AUC and HTER curves under protocol O&C&I to M.
6. Conclusion
SSAN separates content and style features for generalizable face anti-spoofing and applies distinct learning strategies to them. Experiments on existing and proposed benchmarks demonstrate the method's effectiveness.
- Method: SSAN separates content and style features because they have different properties for generalizable face anti-spoofing.The framework operates on global and local statistics separately.
- Method: Adversarial learning makes content features domain-indistinguishable, while contrastive learning emphasizes liveness-related style information and suppresses domain-specific information.Correct content-style pairs are reassembled for living-versus-spoofing classification.
- Evaluation: A large-scale benchmark aggregates existing datasets to bridge the gap between academic and industrial face anti-spoofing evaluation.The conclusion presents the benchmark as part of the proposed solution.
- Results: Results on existing and proposed benchmarks demonstrate the superiority of the proposed methods.The conclusion reports effectiveness across both benchmark types without specifying a numerical result.