submitted on 2024-10-28, 10:01 and posted on 2024-10-30, 10:16authored byWasmiya Abdulla M. S. Aldosari
Federated Learning is an emerging privacy-preserving machine learning framework, mainly based on the federation of multiple devices that a central server coordinates. The central server is jointly trained using all these devices, referred to as clients, without sharing each client’s raw data. What is shared are updates of machine learning parameters computed locally for each client. Eventually, federated learning found its way to authentication, particularly user authentication. Using federated learning for authentication demonstrated a distinctive approach, different from how user authentication models are commonly trained centrally using machine learning models. User authentication in the federated settings enables participants to preserve their information from both the server and other participants, making it hard for adversaries to obtain sensitive information in adversarial environments. However, pairing user authentication systems with federated learning makes them prone to information leakage threats.In this thesis, we propose a federated user authentication with differential privacy system, where local differential privacy is added to local weights before distributing them to the server for the averaging process. we have evaluated our model using root mean squared error, precision, and receiver operating characteristic curve. We also showed that our model resists the membership inference attack, which the plain federated user authentication model is susceptible to as a result of the data leakage problem. Compared with the federated user authentication system without differential privacy, our model minimizes data leakage and successfully lowers the amount of information an adversary can infer about a target client.