PaperlayerView exact revision on arXiv

Source-linked AI summary

NAS-FAS: Static-Dynamic Central Difference Network Search for Face Anti-Spoofing

Zitong Yu, Jun Wan, Yunxiao Qin, Xiaobai Li, Stan Z. Li, Guoying Zhao

arXiv:2011.02062v1cs.CV

TL;DR

NAS-FAS addresses the difficulty of finding FAS architectures that remain robust under domain shifts and unseen spoofing attacks. It searches task-aware spaces using central-difference operators, static-dynamic representations, and Domain/Type-aware Meta-NAS, and achieves state-of-the-art performance across nine benchmark datasets and four protocols. The authors also establish cross-dataset cross-type evaluation with CASIA-SURF 3DMask, while noting remaining performance gaps in some settings.

  • Problem

    FAS architecture search must handle domain shifts and unseen spoofing attacks, while expert-designed networks may be sub-optimal for the task.

  • Method

    NAS-FAS searches a task-aware space combining central-difference convolution and pooling, static-dynamic representations, and Domain/Type-aware Meta-NAS.

  • Results

    NAS-FAS achieves state-of-the-art performance on nine FAS benchmark datasets with four testing protocols.

  • Takeaways & Limitations

    Cross-dataset cross-type evaluation with CASIA-SURF 3DMask assesses NAS transfer and generalization across unseen domains and spoofing attack types.

  • Takeaways & Limitations

    Performance remains below practical requirements in some settings, including less than 70% AUC and less than 20% HTER on 3DMask.

Abstract

from arXiv · show

Face anti-spoofing (FAS) plays a vital role in securing face recognition systems. Existing methods heavily rely on the expert-designed networks, which may lead to a sub-optimal solution for FAS task. Here we propose the first FAS method based on neural architecture search (NAS), called NAS-FAS, to discover the well-suited task-aware networks. Unlike previous NAS works mainly focus on developing efficient search strategies in generic object classification, we pay more attention to study the search spaces for FAS task. The challenges of utilizing NAS for FAS are in two folds: the networks searched on 1) a specific acquisition condition might perform poorly in unseen conditions, and 2) particular spoofing attacks might generalize badly for unseen attacks. To overcome these two issues, we develop a novel search space consisting of central difference convolution and pooling operators. Moreover, an efficient static-dynamic representation is exploited for fully mining the FAS-aware spatio-temporal discrepancy. Besides, we propose Domain/Type-aware Meta-NAS, which leverages cross-domain/type knowledge for robust searching. Finally, in order to evaluate the NAS transferability for cross datasets and unknown attack types, we release a large-scale 3D mask dataset, namely CASIA-SURF 3DMask, for supporting the new 'cross-dataset cross-type' testing protocol. Experiments demonstrate that the proposed NAS-FAS achieves state-of-the-art performance on nine FAS benchmark datasets with four testing protocols.

1 INTRODUCTION

NAS-FAS addresses domain shift and unseen spoof attacks in face anti-spoofing by searching task-aware architectures built around central-difference operators and static-dynamic representations. The method introduces cross-domain/type search and evaluates transferability with new testing resources, reporting state-of-the-art results across nine datasets.

  • Face anti-spoofing is needed because face recognition systems remain vulnerable to print, replay, and 3D-mask presentation attacks.
  • Expert-designed generic classification backbones may lack FAS task-oriented knowledge, while domain shifts and unseen attacks complicate robust architecture search.Traditional cross-entropy can encourage arbitrary patterns such as screen bezels instead of essential spoof cues.
  • Central Difference Convolution and Central Difference Pooling form a FAS-oriented search-space family for fine-grained, invariant spoof representations.The paper motivates these operators with consistent spoofing patterns observed across shifted illumination and camera domains.
  • NAS-FAS combines static-dynamic representation with Domain/Type-aware Meta-NAS to exploit spatio-temporal discrepancy and cross-domain/type knowledge during search.The static-dynamic design targets differences between live and spoofing faces, including dynamic texture, temporal depth, and motion blurriness.
  • The paper establishes the first FAS cross-dataset cross-type testing protocol and releases the large-scale CASIA-SURF 3DMask dataset to support it.The protocol evaluates NAS transfer and generalization across unseen domains and spoofing attack types.
  • NAS-FAS achieves state-of-the-art performance on nine FAS benchmark datasets under four testing protocols.The paper describes these experiments as evaluating the proposed models across the benchmark suite.

2 RELATED WORK

Prior FAS research combines handcrafted or learned features with generalization strategies, but often overlooks task-specific architectures. NAS-FAS addresses this gap with central-difference operators and domain/type-aware architecture search for robust spoof-pattern representation.

  • Neural Architecture Search: NAS has been applied to several computer-vision tasks, but NAS-FAS is presented as the first detailed NAS study for FAS.The paper distinguishes FAS from generic classification because intrinsic live/spoof cues are vulnerable to domain shifts and unknown attacks.
  • Static-Dynamic Face Anti-Spoofing: Multi-frame dynamic FAS methods can be more robust than single-frame methods but require greater communication bandwidth and memory during deployment.The paper proposes a compact static-dynamic representation without extra inference cost.
  • Open-Set Face Anti-Spoofing: Existing FAS models can overfit common domains and attacks, motivating domain-generalization methods such as adversarial, meta-learning, and disentangled learning approaches.These methods improve generalization capacity but do not explicitly address detailed intrinsic spoofing patterns.
  • Open-Set Face Anti-Spoofing: The paper searches architectures that represent discriminative and generalizable spoofing patterns, rather than focusing only on learning strategies.Lattice artifacts are given as an example of an intrinsic spoofing pattern.
  • Convolution and Pooling Operators: Central difference convolution and pooling are introduced because existing operators may have limited capacity for intrinsic spoofing features under domain shift.The resulting search space combines these operators with NAS for robust feature learning and liveness discrimination.

3 STATIC-DYNAMIC CENTRAL DIFFERENCE NET-

The paper builds FAS-aware networks from central-difference convolution and pooling, then combines static appearance with dynamic cues. These components are integrated into depth-supervised central-difference networks for detailed and robust liveness representation.

  • Central Difference Convolution: Central difference convolution replaces weighted local aggregation with center-oriented gradient aggregation to preserve detailed intrinsic features.It is motivated by local binary pattern-style central differences and is intended to improve representation and generalization.
  • Central Difference Convolution: CDC combines intensity-level and gradient-level information, with θ controlling their contributions without adding parameters because weights are shared.When θ=0, CDC reduces to vanilla convolution.
  • Central Difference Pooling: Central difference pooling extends average aggregation with gradient information because uniform average pooling can blur discriminative intrinsic spoofing features.CDP uses λ to adjust the balance between intensity-level and gradient-level information, and average pooling is recovered when λ=0.
  • Central Difference Networks: CDN integrates CDC and CDP into a depth-supervised DepthNet-style backbone to estimate facial depth from static or static-dynamic inputs.Replacing convolutions with CDC yields CDN CDC, while replacing max pooling with CDP yields CDN CDP.
  • Training Design: The default operator trade-off settings are θ=0.7 and λ=0.7, and training uses MSE plus contrastive depth loss for pixel-wise supervision.The input is a 256 × 256 × 3 static or static-dynamic representation.
  • Static-Dynamic Representation: Dynamic images expose live/spoof differences such as depth-aware structure, noise, and lattice artifacts that are less apparent in original static images.Combining dynamic patterns with static images provides both appearance and temporal information for FAS.

4 NEURAL SEARCHING FOR FAS

The paper frames NAS for FAS as a task-aware search problem, using differentiable architecture optimization, specialized FAS search spaces, and domain/type-aware meta-learning to improve robustness to shifted domains and unseen attacks.

  • Differentiable Architecture Search: NAS alternately optimizes network weights on support data and architecture parameters on query data, then discretizes the architecture by selecting highest-weight operations.The final discrete architecture uses M=2 for the baseline search space and M=1 for the FAS search space.
  • Search Spaces: The baseline search space explores eight combinations of static or static-dynamic inputs, vanilla or central-difference operators, and cross-entropy or DeepPixel losses.The FAS search space separately explores ten configurations involving max pooling, CDP, attention, and static-dynamic inputs.
  • FAS Search Space: The FAS search space contains low-, mid-, and high-level cells, with pooling or CDP, optional spatial attention, feature concatenation, and a task-specific loss.Its network is inspired by DepthNet and uses one stem and head layer together with multi-level cells.
  • Domain/Type-aware Meta-NAS: Domain/Type-aware NAS partitions source domains or attack types into support and query sets so architecture search targets an unseen domain or type.D/T-Meta-NAS additionally meta-trains weights across shifted tasks before updating the architecture.
  • Domain/Type-aware Meta-NAS: D/T-Meta-NAS updates architectures using domain/type-generalized weights, making the search more stable under shifted discrepancies and more likely to detect unseen spoofing faces.The method iteratively meta-trains weights and updates architectures using domain/type-shifted knowledge.

5 DATASETS AND PROTOCOLS

The paper introduces CASIA-SURF 3DMask to evaluate FAS generalization under realistic 3D-mask attacks, varied environments, cross-dataset conditions, and unseen attack types.

  • CASIA-SURF 3DMask Dataset: The dataset targets realistic in-the-wild 3D-mask generalization, addressing earlier datasets’ low quality, limited subjects or videos, controlled environments, and unrealistic masks.The paper describes it as the first FAS dataset considering outdoor scenes with challenging lighting.
  • CASIA-SURF 3DMask Dataset: CASIA-SURF 3DMask contains 288 live videos from 48 subjects across six acquisition conditions and 864 spoof videos using three mask-decoration cases.The conditions include normal, back-light, front-light, side-light, outdoor shadow, and outdoor sunlight; spoof masks include cases with or without hair and glasses.
  • Data Collection: Videos are captured with Apple, Huawei, and Samsung mobile devices at approximately 30 fps and 1080p resolution, with sequences lasting about 10 seconds.The recorded subjects are Chinese people, including 21 males and 27 females.
  • Testing Protocols: The evaluation uses nine databases and four protocols, with the first two testing domain-shift robustness and the last two testing generalization to unseen attack types.The final protocol combines serious domain shifts with unseen attack types.

6 EXPERIMENTS

Experiments assess central-difference operators, static-dynamic representation, search-space design, and Meta-NAS across FAS protocols and datasets.

  • Experimental Design: The experiments evaluate central-difference components, static-dynamic representation, baseline and FAS search spaces, Meta-NAS, and state-of-the-art performance across nine datasets and four protocols.Protocol-1 on OULU-NPU is used for component and search-space ablations, while cross-dataset testing evaluates D/T-Meta-NAS.
  • Central Difference Ablation: Larger θ and λ improve CDN performance over vanilla convolution at θ=0, ACER=3.8% and average pooling at λ=0, ACER=4.1%.The result is attributed to adding gradient-based local details.
  • Central Difference Ablation: SelfConv has negative effects in most θ settings and improves ACER by only 0.5% at θ=0.2, indicating difficulty learning intrinsic spoofing patterns from arbitrary local relations.This comparison appears in the CDC and CDP ablation study.
  • Central Difference Ablation: CDC outperforms vanilla, LBConv, GaborConv, and SelfConv by more than 2% ACER, while CDP outperforms average pooling by 2.5% ACER and also exceeds max pooling.All convolution comparisons use best hyperparameters and static-dynamic inputs.
  • Interpretation: Central-difference operators represent local intrinsic spoofing patterns and are less affected by external changes such as illumination, supporting robustness under domain shifts.The paper identifies these gradient clues as discriminative for FAS.

6.3 Static vs Static-Dynamic Representation

The static-dynamic representation is evaluated against static and dynamic inputs as part of the searched FAS design.

  • Static vs Static-Dynamic Representation: Using only dynamic information reduces performance because it loses spatial detail, whereas normalized fusion of static and dynamic inputs improves ACER by 0.3% over static RGB inputs.The comparison evaluates static and dynamic inputs from both separation and fusion perspectives.
  • Static vs Static-Dynamic Representation: The paper treats dynamic information as a spatio-temporal difference compatible with CDC- and CDP-based spatial differences in the task-aware search.Dynamic-static representation, CDC, and CDP are jointly included as positive task-aware knowledge.

6.4 NAS with Baseline Search Space

NAS with the baseline search space exposes a gap between generic object-classification search and FAS, while adding FAS-aware operators and representations improves the searched architectures.

  • NAS Gap between Object Classification and FAS tasks: 8.6% ACER: vanilla-convolution NAS with cross-entropy fails to detect spoofing robustly under slight domain shift, versus 5.2% for static ResNet50.The comparison suggests domain shift and a sub-optimal vanilla search space contribute to the NAS gap for FAS.
  • Impact of Task-Aware Knowledge: CDC and CDP improve architectures across tasks, but directly searching on FAS remains necessary for task-aware architecture discovery.The baseline comparison reports that object-classification-searched architectures perform worse than FAS-searched ones under a fair baseline-space comparison.
  • Impact of Task-Aware Knowledge: 0.6% ACER: NAS Van Max outperforms non-NAS DepthNet with static-dynamic representation.The FAS search space incorporates multi-level features, depth-wise supervision, and spatial attention.
  • NAS-FAS Architecture: The default NAS-FAS setting uses static-dynamic inputs with CDC, CDP, and attention, producing a compact architecture with a more complex mid-level cell.The searched network is shown in Fig. 12(c).

6.6 Comparison with Random Sampling

The experiments evaluate PC-DARTS against random sampling across search spaces and domains, then report NAS-FAS performance across standard and cross-type benchmarks.

  • Under Different Search Spaces: RI measures the relative improvement of PC-DARTS over random sampling using ACER on Protocol-1 of OULU-NPU.Four search spaces are compared: Van-Baseline, CD-Baseline, Van-FAS, and CD-FAS.
  • Searching on Multiple Domains: +5.9% RI for Baseline and +2.7% RI for FAS: Domain/Type-aware Meta-NAS improves search across multiple domains.Without meta-updating, D/T-Baseline and D/T-FAS are less robust than their original counterparts.
  • Results on OULU-NPU: 0.2%, 1.3%, 1.8%, and 5.0% ACER: NAS-FAS ranks first on all four OULU-NPU protocols.The searched architecture transfers across external environment, attack-medium, and camera-variation conditions despite being searched on Protocol-1.
  • Results on SiW: 0.12%, 0.04%, and 1.52% ACER: NAS-FAS performs best on all three SiW protocols.It surpasses CDCN++ by 0.38% ACER on the most challenging SiW Protocol-3.
  • Cross-Type Testing: NAS-FAS achieves the best overall performance in cross-type testing on CASIA-MFSD, Replay-Attack, and MSU-MFSD.The result is reported against methods including the zero-shot learning method DTN.

6.8 Cross-Dataset Intra-Type Testing

Cross-dataset intra-type testing treats four datasets as domains and evaluates architectures trained on three domains and tested on the remaining one.

  • Protocol: Three of OULU-NPU, CASIA-MFSD, Replay-Attack, and MSU-MFSD are used for training, while the remaining dataset is reserved for testing.The datasets share attack types but differ in environments, motivating domain-aware Meta-NAS.
  • Results: Architectures searched on OULU-NPU Protocol-1 remain effective under unseen environments, while D-Meta-NAS improves search quality across multiple source domains.The improvement applies to both baseline and FAS search spaces.

6.9 Intra-Dataset Cross-Type Testing

Intra-dataset cross-type experiments test generalization from known to unknown spoofing attacks, including mask-related transfer settings and type-aware meta-search.

  • Cross-Type Testing: NAS-FAS achieves the best overall performance for replay-versus-print cross-type testing on CASIA-MFSD, Replay-Attack, and MSU-MFSD.The authors attribute this to representation of intrinsic spoofing patterns shared by seen and unknown attacks.
  • SiW-M Cross-Type Testing: 24% and 26% improvement: NAS-FAS improves over the previous state of the art in overall ACER and EER on SiW-M.The model detects almost all Impersonation and Partial Paper attacks with EER<1%.
  • Type-Aware Meta-NAS: 4.3% EER reduction for Baseline and 1.8% EER decrease for FAS: T-Meta improves type-shifted search spaces.The result is reported for type-aware meta-search on cross-type testing.
  • Limitations: NAS-FAS with T-Meta performs worse than NAS-Baseline on some Print and Half sub-protocols.The authors suggest biased attention and conflicts during meta-search across multiple type-shifted tasks, and identify more robust unseen-attack NAS as future work.
  • Type-Aware Search Setup: T(Mask)-Meta searches architectures on Replay, Print, Makeup, and Partial attacks while excluding Mask attacks for subsequent cross-dataset cross-type testing.The resulting baseline and FAS architectures are visualized in Fig. 12(b) and (d).

6.10 Cross-Dataset Cross-Type Testing

The cross-dataset cross-type protocol trains on OULU-NPU and SiW, then tests on unseen domains and mask attacks. NAS-FAS generalizes better than several classical models on the challenging 3DMask dataset, but remains below practical requirements.

  • Protocol: Training uses OULU-NPU and SiW, while testing uses 3DMAD, HKBU-MARs, and 3DMask with unseen mask attacks.The protocol combines domain shift with attack-type shift.
  • Results: NAS-FAS achieves less than 20% HTER on 3DMask, but this remains far from the desired level for practical use.The authors present 3DMask and the protocol as resources for future FAS research.
  • Transferability: PC-DARTS with the Baseline search space performs near worst on the three mask datasets, whereas NAS-Baseline and NAS-FAS achieve significantly better performance.The comparison indicates stronger transfer capacity for the task-aware search spaces under domain and attack shifts.

7 CONCLUSION

The paper introduces NAS-FAS as the first NAS approach for FAS, using task-aware operators, static-dynamic representations, and domain/type-aware Meta-NAS. It establishes cross-dataset cross-type evaluation with CASIA-SURF 3DMask, while leaving broader search-space optimization for future work.

  • Conclusion: NAS-FAS is presented as the first neural architecture search approach for face anti-spoofing.The method applies task-aware knowledge to the FAS search space.
  • Conclusion: The approach combines static-dynamic representation, central difference convolution and pooling, and domain/type-aware Meta-NAS.These components target robust architecture discovery across source datasets and attack types.
  • Conclusion: The paper establishes a cross-dataset cross-type protocol with the CASIA-SURF 3DMask dataset to evaluate NAS transferability for FAS.Extensive experiments are used to verify the proposed methods.
  • Future work: Future work includes searching CD-operator parameters across layers and channels and exploring multi-branch networks for FAS.The stated directions include static and dynamic branches and multimodality.
Loading 2011.02062v1…