One-class learning for face anti-spoofing via pseudo-negative sampling

Mi-Gyeong Gwon, Wonjun Kim

Abstract: Face anti-spoofing, which aims to defend against various attacks in face authentication systems, has drawn increasing attention with the advance in biometrics. Although many studies for face anti-spoofing have shown the remarkable performance improvement, they still suffer from a lack of generality, i.e., vulnerability to unknown attacks frequently occurring in real-world scenarios. Recently, the one-class learning approach has emerged as a way to overcome the aforementioned issue thanks to its capability to distinguish unseen forgery types by precisely understanding the subtle pattern of real faces. However, it is quite a difficult problem to determine the accurate decision boundary by only using real facial images. To cope with this limitation, we propose to apply a novel pseudo-negative sampling scheme in one-class learning for face anti-spoofing. More concretely, pseudo-negative samples are generated based on the statistical distribution of real facial samples and utilized as the proxy of fake facial samples to construct the robust decision boundary. The proposed method is designed by following a two-stage unsupervised learning framework. Firstly, the model learns the feature representation of real faces via the Siamese architecture in the pre-training stage. In the fine-tuning stage, pseudo-negative features are randomly sampled at a suitable distance from real facial features in the latent space. These sampled features are then utilized with real facial features to guide the classifier. Since such pseudo-negative features are not limited to specific fake properties, our classifier can effectively learn to distinguish real and fake faces without using any fake facial images during the training. Experimental results on benchmark datasets show that the proposed method is effective for face anti-spoofing even with unseen spoofing attacks, which achieves the state-of-the-art performance on the Replay-Attack dataset, i.e., 94.48% in AUC. The code and model are publicly available at: https://github.com/DCVL-FA/PNS-release.

External IDs:dblp:journals/mta/GwonK24