Searching Central Difference Convolutional Networks for Face Anti-Spoofing (2003.04092v1)

Abstract: Face anti-spoofing (FAS) plays a vital role in face recognition systems. Most state-of-the-art FAS methods 1) rely on stacked convolutions and expert-designed network, which is weak in describing detailed fine-grained information and easily being ineffective when the environment varies (e.g., different illumination), and 2) prefer to use long sequence as input to extract dynamic features, making them difficult to deploy into scenarios which need quick response. Here we propose a novel frame level FAS method based on Central Difference Convolution (CDC), which is able to capture intrinsic detailed patterns via aggregating both intensity and gradient information. A network built with CDC, called the Central Difference Convolutional Network (CDCN), is able to provide more robust modeling capacity than its counterpart built with vanilla convolution. Furthermore, over a specifically designed CDC search space, Neural Architecture Search (NAS) is utilized to discover a more powerful network structure (CDCN++), which can be assembled with Multiscale Attention Fusion Module (MAFM) for further boosting performance. Comprehensive experiments are performed on six benchmark datasets to show that 1) the proposed method not only achieves superior performance on intra-dataset testing (especially 0.2% ACER in Protocol-1 of OULU-NPU dataset), 2) it also generalizes well on cross-dataset testing (particularly 6.5% HTER from CASIA-MFSD to Replay-Attack datasets). The codes are available at \href{https://github.com/ZitongYu/CDCN}{https://github.com/ZitongYu/CDCN}.

• The paper presents Central Difference Convolution (CDC) to fuse intensity and gradient features for enhanced face anti-spoofing.
• It proposes CDCN and CDCN++ architectures that leverage frame-level processing and Neural Architecture Search to eliminate the need for extensive video sequences.
• Comprehensive tests show CDCN++ achieves a 0.2% ACER on OULU-NPU and robust generalization with a 6.5% HTER across diverse datasets.

Overview of "Searching Central Difference Convolutional Networks for Face Anti-Spoofing"

The paper presents a novel approach to face anti-spoofing (FAS) using Central Difference Convolution (CDC) to improve the robustness and accuracy of Convolutional Neural Networks (CNNs) in this domain. Traditional FAS methods either rely heavily on complex network architectures or require extensive input sequences, presenting challenges in capturing fine-grained details and adapting to environmental changes.

Key Contributions

1. Central Difference Convolution (CDC): The authors introduce CDC as a new convolution operator designed to integrate both intensity and gradient information. CDC is positioned as a more capable feature extractor, particularly for invariant features, when compared to vanilla convolution.
2. Central Difference Convolutional Network (CDCN): The paper introduces a CDC-based network, demonstrating superior modeling capabilities for FAS tasks. The architecture leverages a frame-level approach, sidestepping the need for extended video sequences and making it conducive to real-time applications.
3. Neural Architecture Search (NAS): A NAS framework is established over a CDC-centric search space to automate the identification of performant network architectures, culminating in CDCN++ that employs a Multiscale Attention Fusion Module (MAFM).
4. Empirical Validation: Comprehensive testing on six benchmark datasets confirms the efficacy of the approach. CDCN++ achieves a notable 0.2% ACER on the OULU-NPU dataset and demonstrates robust generalization with a 6.5% HTER from CASIA-MFSD to Replay-Attack datasets.

Implications and Future Work

The introduction of CDC offers a promising avenue for enhancing CNN performance in high-stakes security applications where FAS is critical. By focusing on frame-level processing and robust feature extraction, this approach may facilitate real-time deployment in systems where rapid response is vital.

Further refinements could explore adaptive versions of CDC to dynamically adjust the contribution of gradient data, potentially enhancing performance across varying datasets and lighting conditions. Expanding CDC's application beyond FAS to other areas requiring fine-grained feature detection could further demonstrate its utility.

Conclusion

This work contributes significantly to the FAS literature by introducing and validating CDC as a potent tool for enhancing traditional CNN architecture's ability to discern between live and spoofed faces. Through robust empirical testing and the integration of NAS and attention mechanisms, the paper lays a foundation for future enhancements and applications in real-world AI deployments.