FUSE: a federated learning and U-shape split learning-based electricity theft detection framework

In this study, we propose a novel theft detection framework named FUSE. Firstly, we introduce a new variant of split learning named three-tier U-shape split learning into the local training process. This allows us to migrate the extensive computational overhead to the assisted CSs, while ensuring the sensitive data is preserved in the place where it is generated for privacy-preserving. Furthermore, we design a two-stage semi-asynchronous aggregation mechanism to accommodate the straggler issue and associated communication overhead, which consists of cosine similarity-based pre-aggregation and staleness-aware aggregation. Finally, we conduct extensive experiments and validate our model performance through the comparisons with the benchmarks.