Research On Key Technologies Of Continuous Authentication Based On User Behavioral Biometrics

With the continuous development and advancement of mobile technology,the boundaries of mobile smart devices have been broken.As the first line of defense for network information security,authentication is crucial to ensuring users’ safe access to mobile smart devices.However,the existing authentication methods based on knowledge and physiological biometrics are all one-time authentication,which cannot guarantee the legitimacy of the current user identity during the use of mobile intelligent devices.To address these issues,current researchers have conducted extensive research on continuous authentication systems based on user behavioral biometrics,but there are still issues such as insufficient training data for continuous authentication in single-device scenarios,insufficient coordination for continuous authentication among devices in multi-device scenarios,and insufficient use of authentication models with similar behavioral biometrics in heterogeneous multi-device environments.To this end,this paper proposes continuous authentication based on WGAN(Wasserstein Generative Adversarial Networks)augmented sensor data,continuous authentication based on multimodal fusion of heterogeneous device information,and continuous authentication based on hierarchical transfer learning similar behavioral biometrics.The specific research contents are as follows:1.To address the problem of insufficient training data for continuous authentication in single-device scenarios,which leads to poor authentication results,In this paper,we propose a WGAN augmented sensor data for smartphone continuous authentication.This approach augments existing data and improves the accuracy of the authentication model when training data samples are scarce.Specifically,based on sensor data such as accelerometers,gyroscopes,and magnetometers,WGAN is utilized to generate additional data in the training data.Based on the augmentation data,we design a convolutional neural network to extract efficient discriminative features from sensor data,and then uses four classifiers,namely RF(Random Forest),OCSVM(One-class Support Vector Machine),DT(Decision Tree),and KNN(K-Nearest Neighbor),to train and classify the network.Finally,the network is tested on sensor data under different user activities.Experimental results show that the continuous authentication approach based on WGAN data augmentation outperforms the continuous authentication approach based on linear operation data augmentation in terms of both authentication accuracy and equal error rate.2.To address the problem of insufficient continuous authentication coordination among devices in multi-device scenarios.In this paper,we propose a continuous authentication approach based on multi-modal fusion of heterogeneous device information.This approach focuses on multi-device scenarios and collects multi-modal behavioral data of users on devices.For example,keystroke data,mouse movement data,touch screen gestures,and motion sensor data.Based on these data,this paper analyzes,filters,and aggregates behavioral information on devices,constructs multi-modal complementary features,such as joint features and collaborative features,and implements decision-making for feature selection and multi-modal heterogeneous information fusion.Finally,a series of verification experiments were conducted on four datasets of user interactions with multiple devices,demonstrating the effectiveness of collaborative continuous authentication on multiple devices.The experimental results show that compared to the problem of high misidentification rate caused by user behavior changes in multi-device scenarios that can only be addressed by continuous authentication on a single device,this paper utilizes collaboration between multiple devices to improve the limitations of continuous authentication on a single device,resulting in significant improvements in misidentification rate,authentication accuracy,and model scalability.3.To address the problem of how to utilize similar behavioral biometrics authentication models for existing devices in the context of new device migration.In this paper,we propose a continuous authentication based on STL(Stratified Transfer Learning)similar behavioral biometrics.This approach uses user behavioral biometrics data on different devices,utilizes correlation between features,finds similar data for the same user on different types of devices,and then uses MMD(Maximum Mean Discrepancy)distance measurement to measure similar source and target domains,converting them into a new feature space,and based on this,performs intra-class feature migration.Finally,the classifier is trained using the feature vectors generated by intra-class feature transfer,and the click and swipe data on the new device is validated and tested.The experimental results show that compared with traditional dimensionality reduction approachs PCA(Principal Components Analysis)and KPCA(Kernel Principal Component Analysis),STL has an accuracy improvement of 15%to 20%,and compared with transfer learning approachs TCA(Transfer Component Analysis)and TKL(Transfer Kernel Learning),STL has an accuracy improvement of 10%to 15%.The STL-based migration authentication model not only has cross-platform portability,but also enables fast and accurate continuous authentication of new devices.