User Authentication System Based On Insole And Design Of Adversarial Attacks Defense Model

The increasing number of smart devices has become an important tool for users to communicate and interact on the Internet,so the security and privacy of smart devices have become particularly important.The existing user authentication methods(password,fingerprint authentication,face authentication,etc.)are one-time authentication,which can not ensure the legitimacy of users in the process of use.The authentication method based on behavior characteristics can collect the user's behavior data through the built-in sensor of the device,provide continuous and implicit authentication for users,and ensure the legitimacy of users in the process of using the equipment.As a biological feature,foot behavior is more difficult to imitate than fingerprints,faces and iris,which has attracted the attention of researchers.Therefore,this dissertation designs an intelligent insole embedded with pressure and inertial sensors,collects the data of 50 volunteers,including three kinds of daily behavior activities,and makes an exploratory research on the use of foot behavior characteristics for user authentication.A user authentication model based on full convolution neural network is proposed.The error rates of this model in sitting,standing and walking activities are 1.62%,1.71%and 1.73%.respectively.Through the experimental results,the effects of different activity types,different genders,different BMI values,different feet and different types of sensors on the performance of the model are analyzed.Although the deep neural network has strong performance,the research shows that the deep neural network is vulnerable to the attack of countering samples,which leads to the wrong output of the network model.The above defects are a great threat to the solution of using depth model as user authentication.In order to explore the impact of counterattack on the security of user authentication model,this dissertation selects five representative counterattack algorithms to attack the proposed user authentication model,in order to verify the vulnerability of the user authentication model to antagonistic samples.At the same time,a defense model based on generating confrontation network is proposed.The confrontation samples generated by the confrontation attack algorithm are used as the training samples of the defense model,and the attention mechanism and model compression algorithm are added to the defense model.Used to stabilize the training process and reduce the number of parameters in the model.Through the attack algorithm that the user-defined classifier needs to defend,the countermeasure samples are generated to complete the training of the discriminator,and the classifier which can defend against a variety of counter attacks is obtained.The experimental results show that the error rates of the model for the five antagonistic samples are 0.55%,2.69%,1.83%,2.34% and 1.12%,respectively.