AUTHOR=Ma Yukun , Lyu Chengzhen , Li Liangliang , Wei Yajun , Xu Yaowen 

TITLE=Algorithm of face anti-spoofing based on pseudo-negative features generation

JOURNAL=Frontiers in Neuroscience

VOLUME=Volume 18 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1362286

DOI=10.3389/fnins.2024.1362286

ISSN=1662-453X

ABSTRACT=Advanced face anti-spoofing methods have been developed, but most of them are based on collected datasets or currently known attacks. The generalization performance of these learning-based methods against novel attacks needs to be improved. This paper explores the differences in feature distribution between bona fide and attack samples by employing convolutional neural network on existing data.The analysis shows that the distribution of bona fide sample features is relatively clustered, while attack sample features are usually more scattered. The distinction arises from the differences in data collection methods; particularly, attack samples are typically obtained through secondary imaging.Due to their multi-step nature, attack samples are more prone to environmental and device-related variations, resulting in a wider array of influencing factors. Based on these findings, this paper proposes an algorithm to enhance model performance and overcome the limitations associated with incomplete data in practical applications. The algorithm utilizes a convolutional neural network to extract features from existing data and generates pseudo-negative features that align with the characteristics of attack samples in the feature space, thereby assisting the original features in training the model. The experimental results demonstrate that the method proposed in this paper achieves improved performance in both intra-dataset and cross-dataset testing with limited training data.