Abstract:

To enhance the operational security of power systems and ensure the stable functioning of the electricity market, a detection method based on a HeteroFL framework was proposed to address the insufficient capability of electricity theft detection in energy security. Firstly, the approach integrated the CKKS fully homomorphic encryption scheme and data imbalance handling strategies to improve both detection accuracy and data privacy. Secondly, without sharing users' sensitive information, the method built local models for each power retailer and collaboratively trained them to construct a unified electricity theft detection model. Finally, by considering the significant imbalance between normal and malicious electricity consumption data, feature balance and an attention-based prototype mechanism were introduced to enhance the model's ability to identify anomalous patterns in electricity consumption time series. Experimental results indicate that the proposed HeteroFL detection framework, while guaranteeing end-to-end user data privacy, significantly enhances the accuracy and robustness of electricity theft identification. Moreover, it demonstrates superior generalization performance under severe data imbalance conditions, offering an effective and scalable paradigm for other privacy-sensitive applications with pronounced distributional skew.

Key words: electricity theft detection, heterogeneous federated learning (HeteroFL), local model, feature balance, cheon kim kim song (CKKS) homomorphic encryption