Inferring User Input through Smartphone Gyroscope

Currently, smartphones have become an important way for personal privacy leakage. The sensors embedded in smartphones have low-privileged access policy. Attackers could acquire sensor data without authorization. Therefore, we proposed an input recognition attack scheme based on MEMS gyroscope for smartphone keyboards. By collecting and analyzing the gyroscope 3-axis data during the user clicking screen, a deep learning model is established for mapping between the user input and the gyroscope data. We call this model the input recognition model. To improve the attack efficiency of the scheme, we designed a window function to extract the feature information from the Gyro data. The features are converted into binary images, and then the Residual Network is used to train the input recognition model. The accuracy of this model has reached more than 96%, and the recall rate and F1 value are better than others.