AuthConFormer: Sensor-based Continuous Authentication of Smartphone Users Using A Convolutional Transformer

Abstract

To prevent the leakage of personal information on smartphones, a method that can authenticate the smartphone user’s identity throughout the session is essential. Sensor-based continuous authentication is a promising potential approach that uses behavioral biometrics to authenticate smartphone user’s identity. Recent studies on continuous authentication have achieved encouraging progress. However, they suffer two limitations: 1) weak capability of capturing smartphone user’s behavioral patterns from biometric data sequences with time correlation because these methods don’t consider the long-range dependencies between the behavioral biometric data sequences; 2) poor performance under a high authentication frequency. To solve these issues, we present AuthConFormer, a novel continuous authentication system based on a proposed convolutional transformer, which is suitable for running on the mobile phone CPUs.The proposed convolutional transformer applies an inverted residual block instead of linear projection in the pure transformer to perform the query, key and value embeddings, respectively. Compared with the pure transformer, the proposed convolutional transformer requires less computation, and can make full use of the contextual information of 2D feature maps. We perform several experiments to evaluate the performance and effectiveness of the AuthConFormer on three public datasets. Experimental results show that the proposed AuthConFormer can achieve the mean values of 1.06% EER, 1.25% EER and 1.19% EER on three public datasets with a high authentication frequency (one authentication every 0.6s), respectively. Besides, the convolutional transformer model pre-trained on the HMOG dataset exhibits good cross-domain generalization capability on the BrainRun dataset.