SecFD-FL: Frequency-Decoupled Federated Learning with Secure Aggregation for Charging Load Forecasting in Internet of Electric Vehicles

Abstract

Accurate charging load forecasting is essential for the reliability and efficiency of the Internet of Electric Vehicles (IoEV), yet it is challenging on two fronts: charging demand is highly stochastic, and the data collected by distributed sensors is privacy-sensitive. Although federated learning (FL) preserves data privacy, collaborative training still remains vulnerable to an honest-but-curious server that can invert shared model updates to reconstruct private data. To address both data volatility and privacy concerns, we propose SecFD-FL, a frequency-decoupled federated learning framework for IoEV comprising: 1) a dual-stream gated recurrent unit (GRU) network that separately models short-term and long-term charging load volatility, and 2) a lightweight secure aggregation layer for privacy preservation. Specifically, the secure aggregation layer employs a zero-sum additive masking protocol that perturbs each client's upload with high-variance noise, preventing the server from reconstructing private data while preserving the exact aggregate. Extensive experiments have been conducted on both EV-level and charging station-level using real datasets. The experiment results show that SecFD-FL reduces global RMSE by up to 60% over centralized baselines, achieving Jain's fairness up to 0.991 across clients, and reduces DLG-attack reconstruction correlation by 8 times to 15 times relative to FedAvg -- all with no additional per-round communication cost beyond a one-time key exchange.