Source-linked AI summary
Searching Central Difference Convolutional Networks for Face Anti-Spoofing
Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, Guoying Zhao
TL;DR
Face anti-spoofing methods can miss fine-grained spoofing cues under varying illumination, while long-sequence inputs hinder rapid deployment. This paper introduces CDC-based frame-level networks, including NAS-designed CDCN++ with MAFM, and reports state-of-the-art performance across six benchmark datasets for intra- and cross-dataset testing.
Problem
Existing face anti-spoofing methods can inadequately represent fine-grained information under varying illumination, motivating robust integration of local descriptors with convolution.
Method
The paper develops Central Difference Convolution, CDCN, and NAS-designed CDCN++ with a searched backbone and Multiscale Attention Fusion Module for frame-level face anti-spoofing.
Results
The proposed method achieves state-of-the-art performance on all six benchmark datasets in both intra- and cross-dataset testing.
Takeaways & Limitations
CDC provides a plug-and-play alternative to vanilla convolution with more robust modeling capacity, while CDCN++ aggregates multi-level CDC features through MAFM.
Takeaways & Limitations
The study describes convolutions in 2D for simplicity, treating extension to 3D as straightforward.
Abstract
from arXiv · showhide
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}.
1. Introduction
Face anti-spoofing is needed because printed images, videos, and other presentation attacks can fool face recognition. The paper proposes CDC-based frame-level networks and reports strong intra- and cross-dataset performance.
- Presentation attacks such as print, video replay, and 3D masks threaten reliable face-recognition deployment, motivating face anti-spoofing.
- Stacked-convolution FAS methods can miss fine-grained spoofing patterns and become ineffective under environmental variation such as different illumination.
- Long-sequence methods extract dynamic spatio-temporal features but are difficult to deploy when decisions must be made quickly, motivating high-performing frame-level approaches.
- Central Difference Convolution (CDC) aggregates intensity- and gradient-level information to extract fine-grained invariant spoofing patterns in diverse environments.The paper illustrates lattice artifacts as an example of an intrinsic spoofing pattern.
- Neural Architecture Search is applied over a CDC-specific search space to discover frame-level depth-supervised FAS networks, including CDCN++ with multiscale attention fusion.
- The method achieves state-of-the-art performance across six benchmark datasets under both intra- and cross-dataset testing.
2. Related Work
Related work spans handcrafted and deep-learning FAS methods, while NAS has been applied across computer-vision tasks but not previously to face anti-spoofing. This paper searches frame-level CNNs in a CDC-based space for depth-supervised FAS.
- Handcrafted FAS descriptors capture frame-level spoofing cues, while video-level methods model dynamic clues such as micro-motion and eye blinking.
- NAS has been used for face recognition, action recognition, person re-identification, detection, and segmentation, but not previously for face anti-spoofing.
- The paper searches a frame-level CNN over a CDC-specific search space for depth-supervised face anti-spoofing.
3. Methodology
The methodology combines Central Difference Convolution with depth-supervised FAS, neural architecture search, and multiscale attention to model fine-grained information and build frame-level networks.
- Central Difference Convolution: Central Difference Convolution augments vanilla convolution by aggregating intensity-level and center-oriented gradient information for fine-grained FAS features.The hyperparameter θ controls the tradeoff, with larger values emphasizing central-difference gradient information.
- CDCN: CDCN uses CDC within a depth-supervised network that predicts a 32 × 32 grayscale facial depth map from a single 3 × 256 × 256 RGB facial image.Training combines mean square error loss with contrastive depth loss.
- CDCN++: CDCN++ searches varied low-, mid-, and high-level cells over a CDC operation space, replacing the coarsely repeated CDCN backbone design with a NAS-based backbone.Each cell is represented as a directed acyclic graph whose edges select among candidate operations weighted by architecture parameters.
- CDCN++: MAFM refines multilevel features with spatial attention before concatenation, using vanilla convolutions with 7 × 7, 5 × 5, and 3 × 3 kernels for low-, mid-, and high-level features.CDC is excluded from this module because the authors describe its global semantic cognition as limited for spatial attention.
4. Experiments
Experiments across six datasets evaluate CDCN and CDCN++ using intra-dataset, cross-type, and cross-dataset protocols. Results show strong performance, generalization, and benefits from CDC, NAS-based architecture search, and MAFM.
- Datasets and metrics: Experiments use six databases with intra-dataset, cross-type, and cross-dataset testing, reporting ACER, HTER, AUC, APCER, BPCER, and EER.OULU-NPU and SiW support intra testing; CASIA-MFSD, Replay-Attack, and MSU-MFSD support cross testing, while SiW-M supports cross-type testing.
- Ablation study: CDC outperforms vanilla convolution when θ ⩾0.3, achieving 1.0% ACER at θ = 0.7 versus 3.8% ACER for vanilla convolution.Higher θ increases the contribution of gradient-based details.
- Ablation study: CDC outperforms other convolutions by more than 2% ACER, while LBConv exceeds vanilla convolution and GaborConv performs worst.The comparison uses each convolution's best-performing hyperparameters.
- Ablation study: NAS configurations with varied cells and node attention improve search performance over shared cells with the last intermediate node as output.The authors attribute these gains to more flexible cell structures and selecting the most important output node.
- Intra-dataset testing: CDCN++ ranks first on all four OULU-NPU protocols, recording 0.2%, 1.3%, 1.8%, and 5.0% ACER, respectively.It uses frame-level inputs and its NAS-based backbone transfers across all protocols despite being searched on Protocol-1.
- Intra-dataset testing: CDCN++ performs best on all three SiW protocols and achieves the best overall performance in intra-dataset cross-type testing among replay and print attacks.The SiW protocols cover pose and expression, spoof-medium, and cross or unknown presentation-attack variations.
- Cross-dataset testing: CDCN++ achieves 6.5% HTER when trained on CASIA-MFSD and tested on Replay-Attack, outperforming the prior state of the art by 11%.The authors also report superiority over state-of-the-art frame-level methods in the reverse protocol.
5. Conclusions and Future Work
The paper proposes CDC for face anti-spoofing, builds CDCN and CDCN++ with a searched backbone and MAFM, and verifies them through extensive experiments. It identifies adaptive CDC and broader applicability as future directions.
- The paper proposes Central Difference Convolution (CDC) for face anti-spoofing and builds a Central Difference Convolutional Network (CDCN) from it.
- CDCN++ combines a searched CDC backbone with a Multiscale Attention Fusion Module (MAFM).
- Extensive experiments are performed to verify the effectiveness of the proposed methods.
- Future directions include context-aware adaptive CDC, domain generalization, and applications to image quality assessment and Face-Forensics.
A. Derivation and Code of CDC
This section provides the detailed derivation of CDC and its PyTorch implementation.
- The detailed derivation of CDC is provided in Eq. (7), corresponding to Eq. (4) in the draft.
- The PyTorch implementation of CDC is shown in Figure 9.
B. Adaptive θ for CDC
Adaptive CDC replaces manually selected θ with layer-specific learnable weights constrained to [0, 1]. Its learned values vary by feature level, while performance remains comparable to fixed θ = 0.7.
- Adaptive CDC treats θ as data-driven learnable weights for each layer and uses Sigmoid(θ) to constrain outputs to [0, 1].
- Learned θ values are relatively small in the 2nd–4th and 8th–10th layers but larger in the 5th–7th layers.
- 1.8% vs. 1.0% ACER: adaptive CDC achieves comparable Protocol-1 OULU-NPU performance to CDC with constant θ = 0.7.
C. Cross-type Testing on SiW-M
On SiW-M cross-type testing with 13 leave-one-out attacks, CDCN++ improves overall ACER and EER over the previous state of the art and generalizes across attack types.
- 24% and 26%: CDCN++ improves overall ACER and EER over the previous state of the art on SiW-M cross-type testing.
- CDCN++ detects almost all Impersonation and Partial Paper attacks with EER = 0%.
- CDCN++ sharply reduces EER and ACER for HalfMask, SiliconeMask, TransparentMask, and MannequinHead attacks.
D. Feature Visualization
Figure 11 compares living and spoofing faces through RGB images, low-level features without and with CDC, and spatial attention maps. The visualizations show distinct feature and attention patterns, with CDC features capturing spoofing details more readily.
- Feature comparison: Living and spoofing faces produce visibly different low-level features and corresponding spatial attention maps in MAFM.The passage reports clear differences between the feature and attention-map visualizations for the two face types.
- Feature comparison: Spoofing-face activations appear more homogeneous across facial and background regions than living-face activations.This comparison concerns the low-level features shown in the second and third rows of Figure 11.
- CDC features: CDC features are more likely to capture detailed spoofing patterns than features without CDC.
- Cross-type testing: Table 8 evaluates cross-type testing on the SiW-M dataset.
- Figure 11: Figure 11 arranges living-face examples beside spoofing-face examples across RGB images, low-level features, CDC features, and spatial attention maps.The four rows encode RGB images, low-level features without CDC, low-level features with CDC, and low-level spatial attention maps.