A New Privacy Protection Method for Federated Learning in Smart Grids

Jianguo Wei

Abstract: With the development of grid technology, smart meters have become part of people's lives. Compared to conventional meters, smart meters collect more abundant data and provide more intelligent monitoring information. However, in the power grid, data is often scattered across different locations, making it difficult to fully utilize its value. To better utilize the data, we adopt federated learning to aggregate training data and improve the model's performance, in order to plan local electricity scheduling more effectively. Due to concerns about data privacy, differential privacy with added noise is commonly used, but this approach can significantly impact model accuracy. To address these challenges, we propose a Federated Learning method based on Wiener Filtering for Adaptive Differential Privacy (WADP-FL). WADP-FL adaptively adds noise based on the importance of each layer and utilizes Wiener Filtering to maximize data privacy while preserving model accuracy. Through simulation experiments, we demonstrate that WADP-FL can effectively preserve data privacy in testing neural network models using the MNIST, FMNIST, and CIFAR-10 datasets. Compared to common differential privacy-based federated learning approaches, WADP-FL achieves a significantly improved model accuracy of 4.3%, 2.07%, and 1.86% on different datasets, respectively.

Keywords: Federated Learning; Differential Privacy; Smart Grid; WADP-FL