Federated Learning for User Identification based on Accelerometer and Gyroscope Sensors
Federated Learning, Authentification, Identification, Convolutional Neural Networks
A smartphone can seamlessly collect user behavioral data without requiring additional actions or hardware. Integrated sensors, such as touch and gyroscope, are used in an active or continuous user authentication process to continuously monitor the user and capture behavioral (touch patterns, accelerometer) or physiological (fingerprint, face) data as the user naturally interacts with the device. However, it is not recommended due to privacy concerns related to data transfer from multiple users' mobile devices to a server. This paper presents a Federated Learning (FL) approach that defines a user's biometric behavior pattern for continuous user identification and authentication. The study also evaluates the potential of FL in behavioral biometrics, comparing the performance of Convolutional Neural Networks (CNNs) at different epochs using FL and a centralized method, with minimal chances of incorrect predictions in user identification by the gyroscope. In the era of mobile devices with various sensors, IoT, and other data-sensitive devices, a significant amount of data is shared for application purposes. Smartphones can collect and utilize this data in machine learning approaches, but there is a growing demand for enhanced user security. These sensors can collect user biometric behaviors during their interactions, making it possible to develop a behavioral identification system that uses mobile device data without sharing, focusing on data security and user authentication. This dissertation introduces two Federated Learning approaches that utilize gyroscope data and another with gyroscope and accelerometer data as the unique identifier for continuous user identification. The study employs several CNNs to evaluate their performance for IoT devices, including runtime analysis and identification accuracy within 2 seconds, achieved in just eleven rounds with five epochs.