Privacy-preserving lightweight face recognition

Abstract

Face recognition based on deep learning has become one of the mainstream identity authentication technologies. In recent years, many well-developed deep convolutional neural networks have emerged. But most of these network structures are very complicated, the training processes are very difficult and face recognition based on these network structures are too large to apply to mobile terminals. To solve this problem, we propose a lightweight face recognition algorithm: LightFace, which is based on depthwise separable convolution. In the process of training LightFace, the triplet loss algorithm is used to optimize model. However, LightFace requires a large amount of training data and the parameters of this model will record the users' data, which can be recovered by attackers. To address this problem, we propose an applicable approach to providing strong privacy guarantees for LightFace. In our approach, the generated data and additional noise are used to slightly disturb the original data distribution, so that the attackers cannot correctly predict the training data, which can improve the security of the model. Besides, in the process of training LightFace, ensemble learning increases the randomness of the original data distribution and enhances the robustness of the model. Within the differential privacy framework, we analyzed the privacy loss of the algorithm theoretically and conduct experiments on different face recognition datasets to demonstrate the effectiveness of our privacy preservation method. The experiment results show that LightFace with privacy preservation still has good recognition accuracy while protecting data privacy.