A privacy-preserving heterogeneous federated learning framework with class imbalance learning for electricity theft detection

Electricity theft is a critical issue in smart grids, leading to significant financial losses for utilities and compromising the stability and reliability of the power system. Existing centralized methods for electricity theft detection raise privacy and security concerns due to the need for sharing sensitive customer data. To address these challenges, we propose HeteroFL, a novel heterogeneous federated learning framework for privacy-preserving electricity theft detection in smart grids. HeteroFL enables retailers to collaboratively train a global model without sharing their private data, while accounting for the class imbalance problem prevalent in electricity theft datasets. We introduce a data partitioning and aggregation scheme that assigns different weights to classes, ensuring a balanced contribution and representation of each class in the global model. In addition, our framework leverages the CKKS homomorphic encryption scheme to perform secure computations on encrypted parameters and employs a CNN-LSTM model to capture the spatial and temporal dependencies in electricity consumption patterns. We evaluate HeteroFL using a real-world smart grid dataset and demonstrate its effectiveness and efficiency in detecting energy theft. Furthermore, we analyze the robustness and perform ablation studies to validate the framework's stability and identify the contributions of its key components. Although the impact of approximation errors introduced by the CKKS scheme on the CNN-LSTM model's performance requires further investigation, our framework presents a promising solution for privacy-preserving and accurate electricity theft detection in smart grids using heterogeneous federated learning.