Searching Central Difference Convolutional Networks for Face Anti-Spoofing

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 https://github.com/ZitongYu/CDCN.

Look Locally Infer Globally: A Generalizable Face Anti-Spoofing Approach

“…In Table VI, we compare SSR-FCN with prior work. We find that our proposed method achieves significant improvement in comparison to the published results [20] (relative reduction of 14% on the average EER and 3% on the average ACER). Note that the standard deviation across all 13 spoof types is also reduced compared to prior approaches, even though some of them [1,20] utilize auxiliary data such as depth and temporal information.…”

“…• We show that features learned from local face regions have better generalization ability than those learned from the entire face image alone. [20] under the unknown attack setting, and (ii) 40% on known spoofs. In addition, SSR-FCN achieves competitive performance on standard benchmarks on Oulu-NPU [18] dataset and outperforms prevailing methods on crossdataset generalization (CASIA-FASD [13] and Replay-Attack [12]).…”

“…A majority of prior work on CNN-based face anti-spoofing employs architectures that are densely connected with thirteen convolutional layers [1,20,51,52,54]. Even with the placement of skip connections, the number of learnable parameters exceed 2.7M .…”

“…Based on this observation and experimental results, we utilize the 5-layer FCN (see Table II) with approximately 1.5M parameters. A majority of prior studies employ 13 densely connected convolutional layers with trainable parameters exceeding 2.7M [1,20,52,53].…”

Unified Detection of Digital and Physical Face Attacks

“…along with evident cues for detecting them, studies on anti-spoofs are more sophisticated. The associated JointCNN employs either auxiliary cues, such as depth map and heart pulse signals (rPPG) [37,56,67], or a "compactness" loss to prevent overfitting [19,45]. Recently Stehouwer et al [55] attempt to learn a spoof detector from imagery of generic objects and apply it to face anti-spoofing.…”