Source-linked AI summary
Cross-domain Face Presentation Attack Detection via Multi-domain Disentangled Representation Learning
Guoqing Wang, Hu Han, Shiguang Shan, Xilin Chen
TL;DR
Face PAD must generalize beyond the domains and subjects seen during training, but existing representations can entangle PAD cues with subject- or domain-dependent information. The paper addresses this with DR-Net and MD-Net, and reports better cross-domain performance than state-of-the-art methods. The authors also frame the representation assumptions as expectations validated by experiments rather than guarantees.
Problem
Cross-domain PAD remains limited when training and testing distributions differ, especially for unseen spoof attack types, partly because PAD cues may remain entangled with domain-dependent cues.
Method
DR-Net uses generative models to separate PAD and subject-classification features, while MD-Net learns domain-independent PAD features from multiple domains.
Results
The proposed approach achieves better performance than state-of-the-art face PAD methods in cross-domain PAD.
Takeaways & Limitations
The approach targets more subject-independent and domain-independent representations for robust cross-domain face PAD.
Takeaways & Limitations
The representation claims rely on assumptions that the encoders can ideally extract PAD and identity features, with experiments used to validate these assumptions.
Abstract
from arXiv · showhide
Face presentation attack detection (PAD) has been an urgent problem to be solved in the face recognition systems. Conventional approaches usually assume the testing and training are within the same domain; as a result, they may not generalize well into unseen scenarios because the representations learned for PAD may overfit to the subjects in the training set. In light of this, we propose an efficient disentangled representation learning for cross-domain face PAD. Our approach consists of disentangled representation learning (DR-Net) and multi-domain learning (MD-Net). DR-Net learns a pair of encoders via generative models that can disentangle PAD informative features from subject discriminative features. The disentangled features from different domains are fed to MD-Net which learns domain-independent features for the final cross-domain face PAD task. Extensive experiments on several public datasets validate the effectiveness of the proposed approach for cross-domain PAD.
1. Introduction
Face PAD systems face cross-domain threats because attacks and capture conditions vary, while learned features can entangle spoof cues with subject identity. The paper proposes disentangled and multi-domain learning to improve generalization without requiring a huge PAD dataset.
- Motivation: Face recognition systems are vulnerable to photo, replay-video, 3D-mask, and synthesized face presentation attacks.Deep facial manipulation attacks can be produced using generative models such as VAEs and GANs.
- Motivation: Conventional PAD methods perform promisingly within databases but degrade sharply in new application scenarios.Cross-domain difficulty increases when testing attack types differ from those seen during training.
- Problem: PAD generalization remains limited because learned features may fail to disentangle domain-independent PAD cues from domain-dependent cues.Feature visualization also indicates that representations may cluster by individual subjects rather than by live-versus-spoof labels.
- Approach: DR-Net uses generative models to learn separate encoders for subject-classification features and PAD features.The PAD-GAN and ID-GAN discriminators provide task-specific discrimination during representation learning.
- Approach: MD-Net learns domain-independent PAD features from DR-Net’s disentangled representations across multiple domains.The combined design targets more subject-independent and robust cross-domain PAD representations.
- Result: The proposed approach achieves better cross-domain performance than state-of-the-art face PAD methods.The contribution statement reports this outcome without providing a numerical value in the supplied passage.
2. Related Work
Prior face PAD work includes handcrafted, deep-learning, and domain-generalization methods, while disentangled representation learning provides a related strategy for separating factors of variation. The paper situates its DR-Net and MD-Net design within these research directions.
- Face presentation attack detection: Conventional face PAD methods use texture or temporal cues with handcrafted features and traditional classifiers.Examples include Fourier spectra, LBP, LPQ, SURF, HoG, SIFT, and IDA combined with SVM or LDA.
- Face presentation attack detection: Deep learning approaches address overfitting through transfer schemes, auxiliary supervision, or decomposition of spoof faces into live content and spoof noise.These methods include CNN fine-tuning, depth and rPPG supervision, and inverse spoof decomposition.
- Face presentation attack detection: Domain-generalization approaches use unsupervised adaptation, adversarial adaptation, generalized feature spaces, or semantic partitioning of unknown attacks.The cited methods target cross-domain robustness using MMD, adversarial learning, triplet mining, or deep-tree structures.
- Proposed framework: The proposed framework combines DR-Net for disentangled features with MD-Net for domain-independent representations.Face images from different domains are processed through the two modules before cross-domain PAD.
- Disentangled representation learning: GANs and VAEs have increasingly been used to learn disentangled representations, including unsupervised approaches such as InfoGAN.This literature motivates generative modeling as a way to separate latent factors.
3. Proposed Method
The proposed method combines DR-Net, which disentangles PAD and subject-classification features using generative models, with MD-Net, which learns domain-independent features for cross-domain PAD.
- 3.1. Disentangled Representation Learning via DR-Net: DR-Net learns separate PAD and subject-classification representations through PAD-GAN and ID-GAN encoders.PAD-GAN models live-versus-spoof information, while ID-GAN models subject identity information.
- 3.1. Disentangled Representation Learning via DR-Net: PAD-GAN uses a generator and discriminator whose classification layer distinguishes live and spoof faces, whereas ID-GAN uses subject categories.The discriminator also distinguishes real from generated images, and the overall objective combines GAN and classification losses.
- 3.2. Multi-domain Learning via MD-Net: MD-Net combines disentangled features extracted across two source domains into cross-verified concatenated features for subject classification and PAD.The method uses PAD and ID encoders from both domains rather than concatenating only PAD features.
- 3.2. Multi-domain Learning via MD-Net: The concatenated features are designed to retain identity information from one domain and live-or-spoof information from the other domain.These properties support cross-domain feature learning through classification and image-reconstruction branches.
- 3.2. Multi-domain Learning via MD-Net: After cross-verified multi-task learning, the PAD encoders are expected to generate more general features for cross-domain PAD.The learned PAD encoders are then used to obtain PAD features for face images from the source domains during inference.
P AD and EB
The method extracts two PAD features for each face image and concatenates them for live-versus-spoof classification. At inference, the same procedure is applied to images from an unseen domain.
- Each face image is processed by P AD to extract two PAD features.
- The two extracted PAD features are concatenated into one feature for learning the live-versus-spoof classifier FMD.
- For an unseen-domain face image XU, inference extracts two PAD features and classifies their concatenation with FMD.
4. Experimental Results
Experiments evaluate cross-domain PAD across four databases and three training/testing protocols, comparing the proposed approach with conventional, domain-generalized, and adaptation baselines. Results indicate benefits from multi-domain disentangled representation learning, while ablations and visualizations examine component contributions and failure cases.
- Databases and Protocols: Evaluations use Idiap, CASIA, MSU, and OULU as separate domains in cross-domain protocols with varying numbers of source datasets.Protocol I trains on three datasets and tests on the fourth; Protocol II trains on two datasets and tests on another.
- Experimental Setup: The study reports HTER and AUC for cross-domain testing and compares against domain-generalized, domain-adaptation, and conventional PAD baselines.Baselines include Auxiliary-supervision, MMD-AAE, MADDG, MS LBP, CNN, IDA, CT, and LBPTOP.
- Protocol I Results: Under Protocol I, the proposed approach outperforms MADDG and MMD-AAE, while conventional methods generally perform worse under large training–testing distribution disparities.The discussion attributes these disparities particularly to unseen spoof attack types, illumination, and display devices.
- Protocol II Results: Under Protocol II, the proposed approach shows promising performance against conventional methods and MADDG, indicating stronger exploitation of domain-independent features.The authors associate this advantage with generative disentanglement of identity information from face images.
- Protocol III Results: Under Protocol III, PAD-GAN is comparable to domain-adaptation methods and better than CNN in some settings, but adaptation methods perform better in cases involving CASIA.The authors relate the CASIA difficulty to cut photos absent from other datasets.
- Ablation and Visualization Analysis: Ablations show that removing DR-Net, MD-Net, LCE, or LREC reduces performance, with LCE having greater influence than LREC.Visualizations also show PAD-GAN suppresses identity detail more than ID-GAN, while errors involve illumination, color distortion, and blurriness.
5. Conclusions
Cross-domain face PAD is difficult because presentation attacks and environmental factors vary. The proposed DR-Net and MD-Net approach outperforms state-of-the-art methods on several public datasets, though incorrect classifications remain documented.
- Diverse presentation attack types and environmental factors make cross-domain face PAD challenging.
- Incorrect PAD results include spoof faces classified as live and live faces classified as spoof across Idiap, CASIA, MSU, and OULU.
- The approach combines disentangled representation learning in DR-Net with multi-domain feature learning in MD-Net.DR-Net models live and spoof class distributions, while MD-Net learns domain-independent representations for the final task.
- The proposed approach outperforms state-of-the-art methods on several public datasets.