Mobile Passive Authentication through Touchscreen and Background Sensor Data

The security and usability shortcomings of current mobile user authentication systems based on PIN codes, fingerprint, and face recognition are well known. To overcome such limitations, the present work focuses on the comparative analysis of unimodal and multimodal behavioral biometric traits suitable for mobile passive authentication, such as touchscreen data during separate gestures (keystroke, scrolling, drawing a number, tapping on the screen), and background sensor data (accelerometer, gravity sensor, gyroscope, linear accelerometer, magnetometer).This paper carries out a performance evaluation over one of the most complete and challenging databases to date with mobile user interaction data, HuMIdb, with 600 subjects. For each individual modality, we propose a separate RNN (Recurrent Neural Network) trained with semi-hard triplet loss. In addition, we perform the fusion of the different modalities at score level. Our results show that the best performing tasks are keystroke and drawing a number, whereas the most discriminative background sensor is the magnetometer. Additionally, the fusion of modalities is very beneficial, consistently reducing the Equal Error Rates (EER) by half (ranging from 5% to 13% depending on the modality combination).