Source-linked AI summary
Adaptive Normalized Representation Learning for Generalizable Face Anti-Spoofing
Shubao Liu, Ke-Yue Zhang, Taiping Yao, Mingwei Bi, Shouhong Ding, Jilin Li, Feiyue Huang, Lizhuang Ma
TL;DR
Cross-dataset face anti-spoofing remains challenging, and DG methods have paid limited attention to normalization during feature extraction. ANRL adaptively combines BN and IN with domain- and class-based constraints; it outperforms DG-based methods under four test settings.
Problem
DG-based FAS methods align final features but often overlook normalization during feature extraction, despite BN and IN performing differently across unseen target domains.
Source: §1 · p. 2Method
ANRL fuses BN and IN features with sample-wise factors and uses Inter-Domain Compatible and Inter-Class Separable losses to guide AFNM.
Source: §1 · p. 2, §3.1 · p. 3Results
ANRL outperforms DG-based FAS methods under four test settings.
Source: §4.2 · p. 5Takeaways & Limitations
Feature-map visualizations show selected high-weight IN or BN channels focusing on face regions for intrinsic spoof cues that can generalize to unseen target domains.
Source: §5.1 · p. 7Because fake attacks may be inherently scattered, ICS compacts only the relatively stable real class.
Source: §3.3 · p. 4
Abstract
from arXiv · showhide
With various face presentation attacks arising under unseen scenarios, face anti-spoofing (FAS) based on domain generalization (DG) has drawn growing attention due to its robustness. Most existing methods utilize DG frameworks to align the features to seek a compact and generalized feature space. However, little attention has been paid to the feature extraction process for the FAS task, especially the influence of normalization, which also has a great impact on the generalization of the learned representation. To address this issue, we propose a novel perspective of face anti-spoofing that focuses on the normalization selection in the feature extraction process. Concretely, an Adaptive Normalized Representation Learning (ANRL) framework is devised, which adaptively selects feature normalization methods according to the inputs, aiming to learn domain-agnostic and discriminative representation. Moreover, to facilitate the representation learning, Dual Calibration Constraints are designed, including Inter-Domain Compatible loss and Inter-Class Separable loss, which provide a better optimization direction for generalizable representation. Extensive experiments and visualizations are presented to demonstrate the effectiveness of our method against the SOTA competitors.
1 INTRODUCTION
Cross-dataset FAS remains difficult because deep models that perform well within a dataset can degrade on unseen domains. ANRL adapts normalization per sample and uses domain- and class-oriented constraints to guide generalizable representations.
1 INTRODUCTION
- Deep-learning FAS methods can perform well within a dataset yet degrade significantly when tested across datasets because of biased features and domain gaps.The cited work attributes this gap to models fitting training data while ignoring differences between source and unseen target domains.Source: §1 · p. 2
- Prior DG methods align final features but overlook feature extraction, where BN and IN can perform differently across unseen target domains.The introduction reports BN performing better on one target domain and IN on another, relating the difference to their distinct normalization properties.Source: §1 · p. 2
- ANRL estimates sample-wise factors to fuse BN and IN features, while Inter-Domain Compatible and Inter-Class Separable losses guide factor learning.The losses respectively align feature distributions across domains and enlarge the margin between real and fake samples; meta-learning updates AFNM.Source: §1 · p. 2
- The authors report that experiments and visualizations demonstrate ANRL's effectiveness against state-of-the-art competitors.The introduction gives no specific metric or benchmark values for this comparison.Source: §1 · p. 2
2 RELATED WORK
Related FAS work progresses from handcrafted features and CNN classifiers to domain-generalization methods, but cross-dataset performance remains challenging. Prior normalization combinations use shared parameters, while ANRL's framework adaptively fuses BN and IN features and applies dual constraints.
2.1 Face Anti-Spoofing
- FAS research spans handcrafted descriptors, CNN-based classifiers, auxiliary supervision, specialized operators, and DG approaches, yet cross-dataset performance remains a challenge.The related-work review notes that DG methods align samples in different ways but overlook feature extraction.Source: §2.1 · p. 2
2.2 Batch-Instance Normalization
- Existing BN–IN combinations either replace or supplement BN, or learn shared balance parameters that may not suit sample differences on unseen targets.The review groups these approaches as non-parametric [13] and parametric [7] methods.Source: §2.2 · p. 2
- ANRL's AFNM fuses BN- and IN-normalized features, while Dual Calibrated Constraints guide meta-learning updates toward domain-agnostic, discriminative FAS representations.The framework description says the constraints aggregate same-class source samples and separate real from fake samples.Source: Figure 2 · p. 3
3 PROPOSED METHOD
ANRL combines BN and IN features with sample-specific factors and uses domain- and class-based constraints to guide AFNM. Meta-learning updates AFNM using source-domain splits, while the base model follows normal training.
3.1 Overview
- ANRL combines BN's discriminative features and IN's domain-variation removal, using DCC to guide AFNM toward compact, distinguishing representations.IDC reduces inter-domain distances while scattering samples within a domain; ICS gathers same-class samples and separates different classes.Source: §3.1 · p. 3
3.2 Adaptive Feature Normalization Module
- AFNM estimates sample-specific balance factors from input features and uses them to fuse BN- and IN-normalized representations.It derives channel statistics by global average pooling, then uses attention across channels to select useful information from each normalization.Source: §3.2 · pp. 3–4
3.3 Dual Calibrated Constraints
- DCC combines IDC, which narrows inter-domain gaps and scatters same-domain samples, with ICS, which gathers same-class samples and separates real from fake.Because fake attacks may be inherently scattered, the method compacts only the relatively stable real class.Source: §3.3 · p. 4
3.4 Training Strategy
- Meta-learning splits source domains into meta-train and meta-test sets to simulate domain shifts, updating AFNM while the feature extractor, classifier, and depth estimator train normally.The classifier output is the sole evaluation metric described in the training strategy.Source: §3.4 · pp. 4–5
4 EXPERIMENTS
Experiments evaluate ANRL on four cross-dataset face anti-spoofing tasks, with the target dataset unavailable during training. ANRL outperforms the compared domain-generalization methods across the four settings, and ablations examine its components and normalization choices.
4.1 Experimental Settings
- Evaluation uses four tasks, each training on three of OULU, CASIA, Idiap, and MSU while withholding the fourth dataset as the target.The target dataset is unavailable during training; the tasks are O&C&I to M, O&M&I to C, O&C&M to I, and I&C&M to O.Source: §4.1 · p. 5
4.2 Experimental Results
- ANRL outperforms the compared domain-generalization methods across all four testing settings.The evaluation compares methods on four cross-dataset tasks; Table 1 and Figure 3 present the comparisons and ROC curves.Source: §4.2 · p. 5, Table 1 · p. 6, Figure 3 · p. 6
- With only two source datasets, ANRL achieves the best performance when testing on the remaining CASIA and OULU datasets.MSU and Idiap are used as sources; CASIA and OULU are held out as targets.Source: §4.2 · p. 6, Table 2 · p. 6
4.3 Ablation Study
- On I&C&M to O, ablations report improved performance from meta-learning, domain-compatible alignment, and inter-class separation, with ANRL combining the components performing best.Table 3 evaluates the different components on this task.Source: §4.3 · p. 6, Table 3 · p. 7
- On I&C&M to O, AFNM's sample-specific fusion of IN and BN outperforms fixed or shared-factor alternatives, and adding DCC yields the best results.The reported alternatives include IN-BN-half, BIN, and IBN; Figure 4 visualizes feature representations from variants.Source: §4.3 · p. 7, Table 4 · p. 7, Figure 4 · p. 7
5 ANALYSIS
Feature-map visualizations show that IN and BN emphasize complementary facial cues, while low-weight channels attend to source-specific regions. Balance-factor analyses show that normalization preferences vary by layer, task, and sample.
5.1 Feature Map Analysis
- AFNM combines complementary BN- and IN-weighted channels, which together focus more completely on facial areas for generalization.High-weight channels from either normalization focus on face regions; low-weight channels attend more to nontransferable source-specific cues such as hands and background.Source: §5.1 · pp. 7–8
5.2 Balance Factor Analysis
- Shallow-layer balance factors tend toward zero and high-layer factors toward one, indicating greater IN use for low-level features and BN use for high-level features.The reported pattern is consistent with IN mitigating domain discrepancy in low-level features and BN serving high-level classification features.Source: §5.2 · p. 8, Figure 7 · p. 8
- Balance factors vary across samples, and fake samples show greater variance than real samples, supporting sample-wise selection of normalization factors.The analysis examines channels with large balance-factor variance in layer 1.Source: §5.2 · p. 8, Figure 7 · p. 8
- The authors report smaller feature differences between OULU and CASIA than between Idiap and MSU, and speculate that tasks including Idiap and MSU as source domains use more IN.The t-SNE visualization uses features extracted by an ImageNet-pretrained ResNet.Source: §5.2 · p. 8, Figure 8 · p. 8
6 CONCLUSION
ANRL adaptively combines BN and IN features to learn domain-agnostic, discriminative FAS representations. The authors report that public-dataset experiments demonstrate the method's effectiveness.
6 CONCLUSION
- ANRL uses sample-customized BN and IN feature combinations, guided by domain- and class-oriented Dual Calibration Constraints through meta-learning.The constraints comprise Inter-Domain Compatible Loss and Inter-Class Separable Loss.Source: §6 · p. 8