Research On User Privacy Information Detection Based On Mobile Device Sensor

In recent years,smartphones have rapidly gained popularity with their powerful processing power and diverse applications.The construction of mobile Internet,the rapid development of machine learning and deep learning have further deepened users'dependence on smartphones,and it is gradually difficult for people to leave smartphones in their daily lives.When a user uses a mobile terminal,a large amount of personal privacy information is also generated.Once the information is leaked,it will cause huge losses.Generally speaking,the more types of sensors built into a mobile phone,the more sensitive it is to the perception of the environment,which greatly improves the user experience and increases the risk of user privacy leakage.Earlier,researchers conducted a lot of analytical research on detecting user privacy based on the rich sensors built into smartphones.For example,using the accelerometer and gyroscope to learn the user's click and gesture input;using the magnetic sensor built in the smartphone to infer the user activity of the nearby laptop.However,the existing research has the following shortcomings:the signal is weak,and the noise and enhanced data need to be processed in the later stage,or the attack conditions are harsh,and the device must be close to the victim.Based on many relevant literature and research work at home and abroad,thesis designs two algorithms based on the built-in sensors of smart phones to detect user privacy information.The first algorithm is based on magnetometer to identify and classify applications and in-application services.The second algorithm adds accelerometer and gyroscope data on the basis of the first algorithm for user identification algorithm.Specifically,the main work of this paper is as follows:In the task of identifying and classifying applications and in-application services based on magnetometers,thesis collects magnetometer data generated when users use applications,and uses random forest algorithm after data processing,fourier transform and principal component analysis to extract features.For application identification,use convolutional neural networks for in-app service classification.Compared with other related studies,the segmentation method proposed in this paper can effectively reduce the time and space cost of post-training.It is shown that the scheme in thesis achieves 92%/91%recognition accuracy/macro-averaged score of applications and in-app services.In the user identification and classification algorithm based on accelerometer and gyroscope,based on the recognition and classification task in the previous step,this part adds the accelerometer and gyroscope data generated when the user uses the application,and uses the fully connected network and long short-term memory.After the network extracts features,the task of user identification and classification is performed.Compared with other related researches,thesis applies wavelet multi-resolution analysis to eliminate the influence of user actions,and proposes an embedding vector generation algorithm that can represent the changing trend of flow sequences.It uses fully connected network and long short-term memory network to extract time-frequency features,which effectively improves the expressiveness and accuracy of features.Experiments show that the scheme in this paper achieves a user identification accuracy rate of 92%.