Privacy-preserving federated learning with differential privacy for secure distributed intelligent computing

Abstract Federated Learning (FL) enables collaborative model training without centralized data sharing, making it suitable for privacy-sensitive intelligent computing environments. However, FL models remain vulnerable to information leakage through model updates and prediction outputs, particularly when learning from electroencephalography (EEG) signals. Although Differential Privacy (DP) can mitigate such risks, integrating DP into distributed deep learning often reduces predictive performance and destabilizes optimization. To address these challenges, this study proposes SAFE (Secure Adaptive Federated EEG Framework), a privacy-preserving distributed learning framework integrating representation-level DP, TLS-secured communication, and multi-round Federated Averaging (FedAvg) optimization. SAFE applies Laplace-based DP directly to normalized EEG representations prior to local training while preserving decentralized data locality. The framework was evaluated using EEGNet, ShallowNet, and EEGTCNet on the BCI Competition IV Dataset 2a under multiple privacy budgets ( $$\epsilon = 1, 5, 10$$ ) across 20 communication rounds. Experimental evaluation using classification accuracy, convergence analysis, communication overhead, and Membership Inference Attacks (MIA) demonstrated that SAFE effectively balanced predictive utility and privacy protection. Among the evaluated models, ShallowNet achieved the strongest federated performance and convergence stability. The findings demonstrate the feasibility of integrating DP and secure communication into federated EEG learning without substantial degradation in predictive utility.