Research On Continuous Authentication Of Mobile Terminals For Walking Scene

The existing one-time authentication methods used by mobile terminals mostly use fingerprints,faces,passwords,etc.to confirm the user’s identity when logging in,ignoring the protection of user privacy during the use of equipments.In addition,the collection of physiological characteristics and the input of passwords require frequent cooperation from users,so it is difficult to balance device security and user-friendliness.The continuous authentication method collects user behavior characteristics through the built-in sensor of the mobile device,and can perform frequent implicit authentication in a way that the user does not perceive it.Thus,it is very suitable as the second line of security to make up for the shortcomings of the one-time authentication method.As one of the most common activities in life,gait behavior has broad application prospects in the field of authentication based on behavior characteristics.However,existing continuous authentication schemes based on gait characteristics are mostly designed on datasets collected in controlled environment,so the scene applicability of these schemes is not satisfactory.In response to this problem,we analyze multi-scene walking data,and propose two continuous authentication schemes,which can meet the verification and identification requirements separately.We collect multi-scene walking data and analyze its feature distribution.Combining visual display and numerical analysis methods to explore the influence of time,equipment placement and walking style on the characteristic distribution of gait data.The experimental results show that the users’ gait feature distribution varies from person to person.Among them,the walking style has the greatest impact on the feature distribution,followed by device placement,and time has the least influence.These three factors make the intra-class distance of data expand and the inter-class distance shrinks,which lead to a decrease in the distinguishability among different users and bring more difficulties to authentication.On the basis of data analysis,we propose a continuous authentication scheme based on sparse representation to meet the needs of identity verification.With the goal of balancing the quality of collected data and user experience,we design a non-perceptual data collection method to reduce the consumption of equipment resources by invalid collection.Subsequently,a PCC-based adaptive gait cycle segmentation method is designed to further screen out abnormal data while segmenting the gait sequence.In order to ensure the feature extraction ability of our scheme,we use dictionary learning algorithm to construct a completely data-driven authentication model,and realize the identity authentication function through the weighted fusion sparse representation algorithm.Finally,the performance of the scheme is evaluated through experiments.The scheme can achieve a certification accuracy rate of 93.4% on real walking data,which is an improvement of more than 30% compared with the authentication scheme based on feature engineering.Subsequently,a continuous authentication scheme based on the fusion neural network is proposed to meet the needs of identity recognition.We use a data-driven dictionary learning algorithm to reconstruct gait data,which can achieve preliminary feature extraction while filtering.Aiming at the problem of low discriminability of multi-scene walking data features,a fusion neural network is designed,and the central loss is introduced to define joint loss function,for the purpose of guiding the network training.So that the network can obtain the feature extraction ability while improving its grasp of the recognition ability.Finally,we evaluate the performance of our scheme through experiments,and compare it with related works in this field on the public datasets and the real walking data collected in this article.Experiments show that the use of the joint loss function can improve the recognition ability of the network to a certain extent.Our solution has achieved 99.4% certification accuracy on the ZJU-Gait Acc dataset.And 94.58% and 94.62% authentication accuracy on the Mobi Act dataset and the real walking dataset,respectively,which proves that our solution has good applicability to the multi-scene.