Sensor-Based Continuous Authentication Using Cost-Effective Kernel Ridge Regression

People prefer to store important, private, and sensitive information on smartphones for convenient storage and fast access, such as photos and emails. To prevent information leakage and smartphone illegal access, we propose a novel sensor-based continuous authentication system, SensorCA, for continuously monitoring users’ behavior patterns, by leveraging the accelerometer, gyroscope, and magnetometer ubiquitously built-in smartphones. We are among the first to exploit the data augmentation approach of the rotation, which creates additional data by applying it on the collected raw data and improves the robustness of the proposed system. With the augmented data, SensorCA extracts sensor-based features in both time and frequency domains within a time window, then utilizes the kernel ridge regression with truncated Gaussian radial basis function kernel (KRR-TRBF) to train the classifier, and finally authenticates the current user as a legitimate user or an impostor. We evaluate the authentication performance of SensorCA in terms of different classifiers including KRR-TRBF, KRR-POLY, and SVM-RBF, and the data augmentation approach rotation on KRR-TRBF6 and SVM-RBF. The experimental results show that under the KRR-TRBF6 classifier, SensorCA reaches the lowest median equal error rate of 3.0% with dataset size 8000 and consumes the shortest training time of 0.054 seconds with dataset size 1000.