Research On Continuous Authentication Based On Behavioral Biometrics

With the rapid development of mobile terminal technology,the privacy protection of smartphones has become increasingly critical.Identity authentication for users is one of the important means to protect users' privacy.However,existing authentication approaches(such as passwords,graphical patterns,fingerprint authentication,face authentication)belong to one-time user authentication mechanisms,which authenticate users only once at the time of initial login.Therefore unauthorized users are easily able to physically gain access to unattended smartphones after initial authentication has been performed,which can incur security issues.The authentication method based on behavioral biometrics can automatically collect the data of users' behavioral biometrics by utilizing the built-in sensors of smartphones,thus providing users with continuous and non-perceptual authentication,and ultimately ensuring the legitimacy of the identity of the mobile user in the use of the smartphones.In this paper,we investigate identity authentication on smartphones.Based on the behavioral biometrics of the user,two continuous authentication schemes are proposed.The main research work consists of two parts:1.Continuous Authentication Using Data Augmentation.Continuous authentication on local devices often results in poor performance because there is not enough time to collect a large amount of training data to train the authentication model.This paper breaks through the dependence of traditional authentication schemes on a large amount of training data,and attempts to utilize data augmentation methods to create additionally artificial data.As a consequence,it can increase the amount of the training data and indirectly reduce the time to collect training data.In this paper,we present a novel data augmentation-based continuous authentication system on the basis of users' behavioral patterns.More specifically,it consists of data collection,data augmentation,feature extraction,classifier and authentication.Data collection captures behavioral patterns of users by leveraging two sensors of accelerometer and gyroscope ubiquitously built into smartphones.We are among the first to exploit five data augmentation approaches to create additional data by applying them on training data.Then,sensor-based features are extracted in both time and frequency domains from the augmented data within a time window.We use the one-class support vector machine(SVM)to train the classifier.With the trained classifier and testing features,the system classifies the current user as a legitimate user or an impostor.We evaluate the authentication performance of the system in terms of the impact of window size,accuracy on each of the data augmentation approaches,and accuracy on the combinations of data augmentation approaches,time efficiency,and comparisons with the representative classifiers,respectively.The experimental results show that our authentication system performs highly accurate and consumes shorter authentication time,compared with the authentication methods without data augmentation.2.Continuous Authentication via Two-stream Convolutional Neural Networks(CNN).In the face of complex behavioral biometrics of smartphone users,manually designed features are often time consuming.In this paper,we propose a lightweight two-stream convolution neural network to automatically extract the behavioral biometrics of smartphone users and it is constructed on the basis of deep separable convolution.Furthermore,we propose a novel two-stream convolutional neural network based authentication system.It consists of five modules of the data collection,data preprocessing,feature extraction,classifier and authentication.Data collection captures users' behavioral patterns during smartphone usage,by utilizing the two sensors of the accelerometer and gyroscope omnipresently built-in smartphones.Data preprocessing converts the raw sensor data into two stream of inputs for the CNN.In the feature extraction module,we specially design a two-stream CNN to learn and extract the effective and efficient data representations for resource-constrained mobile devices.The most discriminable ones among these features are extracted by the two-stream CNN.Then,we use the one-class SVM to train the classifier.With the trained classifier and testing features,our system classifies the current user as a legitimate user or an impostor.We evaluate the effectiveness of the two-stream CNN in terms of the accuracy,model parameters,and computational cost,and then evaluate the performance of our authentication system with respect to the equal error rate,time efficiency and comparisons with the representative classifiers,respectively.The experimental results show that the two-stream CNN achieves best performance in comparison with other mobile neural network frameworks,and continuous authentication based on two-stream CNN reaches highly accurate and consumes shorter authentication time,compared with the authentication approaches using manually designed features.