Towards Privacy-Aware Federated Learning for User-Sensitive Data
Muhammad Fudholi Hamidul As’ad, Safa Otoum · 2023
Federated Learning (FL) has been envisioned as a promising approach for collaboratively training learning models while preserving private individuals' data. In the FL training procedure, participants train a global model by exchanging the model parameters and keeping the raw data private. Nevertheless, exchanging those model parameters causes insecure interaction among participants that might disclose the individual's identity or private information. To this end, several approaches have considered secure multiparty computation (SMC) and differential privacy. Those approaches suffer from several drawbacks, such as limited accuracy, computational capacities, or functional behavior, and cannot guarantee participants' identity during the learning process. To this end, in this paper, we propose a novel Threshold Signature-based Authentication (TSA) scheme for secure FL. The TSA scheme secures the participants' identity against any chosen cipher-text attack and forbids external adversaries from malicious attacks. Moreover, the TSA scheme can successfully defend the identity leaks from the trained models against property and membership inference attacks. The experimental results show that the TSA can achieve 91% training accuracy, which is superior to the existing methods.