Android-GAN: Defending against android pattern attacks using multi-modal generative network as anomaly detector

Abstract

Android pattern lock system is a popular form of user authentication extensively used in mobile phones today. However, it is vulnerable to potential security attacks such as shoulder surfing, camera attack and smudge attack. This study proposes a new kind of authentication system based on a generative deep neural network that can defend any attacks by imposters except a registered user. This network adopts the anomaly detection paradigm where only normal data is used while training the network. For this purpose, we utilize both Generative Adversarial Networks as an anomaly detector and Long Short Term Memory that processes 1D time varying signals converted from 2D Android patterns. To handle the stability problem of GANs during the training, Replay Buffer, which has been effectively used in Deep Q-Networks, is also utilized. Evaluation of the proposed method was carried out thoroughly and the accuracy reached to 0.95 in term of the Area Under Curve. Although training this network requires extensive computing resources, it runs on a mobile phone well since the testing version is very light. Further experiments conducted using a group of mobile phone users, including posture variation study, provided comparable performance as well. Results suggest that the proposed system has a potential for real world application.