NAS-FAS: Static-Dynamic Central Difference Network Search for Face Anti-Spoofing (2011.02062v1)

Abstract: 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.

• The paper introduces NAS-FAS, a novel neural architecture search method employing Central Difference Operators and static-dynamic representation to build robust face anti-spoofing systems.
• Novel Central Difference Convolution and Pooling operators are introduced, which enhance the capture of intrinsic spoofing features and improve robustness against domain shifts.
• The method is rigorously evaluated across nine datasets and four protocols, including a cross-dataset cross-type protocol supported by a new CASIA-SURF 3DMask dataset.

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

The paper presents NAS-FAS, a pioneering approach to enhancing face anti-spoofing (FAS) systems using neural architecture search (NAS). The FAS task is crucial for safeguarding face recognition systems from presentation attacks like print, replay, and 3D mask assaults. The proposed method diverges from traditional expert-designed networks, often yielding suboptimal solutions by integrating NAS techniques tailored specifically for FAS. The initiative underlines the importance of identifying well-suited task-aware networks rather than applying generic NAS methods that focus solely on efficiency or classification tasks.

Key Contributions and Findings

1. Search Space Introduction: The paper introduces novel convolution and pooling operators—Central Difference Convolution (CDC) and Central Difference Pooling (CDP)—which enhance the capture of intrinsic spoofing features necessary for discriminative and robust representation. These operators are pivotal in studying the NAS search spaces unique to the FAS task and are shown to perform better in the presence of domain shifts and unseen attack types.
2. Static-Dynamic Representation: This concept introduces a compact representation scheme that fuses static spatial information with dynamic temporal clues. The method uniquely integrates temporal dynamics seen in spoofing actions, thereby enriching the representation used for detection without incurring higher inference costs.
3. Domain/Type-aware Meta-NAS: This approach leverages cross-domain and attack-type knowledge to refine the search for robust architectures that generalize well across unseen situations. The paper reveals how NAS can focus through meta-learning techniques to adapt architectures in the face of variable domains and attack scenarios.
4. Evaluation and Dataset Release: The proposal includes rigorous evaluation across nine benchmark datasets under four testing protocols, including a novel cross-dataset cross-type protocol backed by a newly released CASIA-SURF 3DMask dataset. This dataset is a significant asset for future research, simulating real-world scenarios with high fidelity to practical challenges in FAS.

Implications and Future Directions

The paper stands to impact future development in NAS tailored to specific tasks like FAS, suggesting that task-aware elements such as novel operators and representation strategies are essential for success. The methodology’s potential for transferability to other security domains and challenges associated with domain generalization underscores the relevance of adapting NAS approaches to specialized requirements.

Moreover, the insights from this research signal important avenues for the future, such as exploring the optimal configurations of CDC/CDP parameters, multi-branch network searches in dynamic/static modalities, and constraint-based searches that align with resource limitations.

Overall, NAS-FAS emerges as a refined method that maximizes FAS capabilities, revealing critical opportunities to enhance intelligent recognition systems' robustness and reliability against sophisticated spoofing threats through tailored NAS strategies. The release of a substantial dataset also supports ongoing innovation and validation of anti-spoofing advancements, establishing a benchmark for evaluating system efficacy in complex and varied conditions.