Abstract: Electricity theft is one of the major challenges in the new type power systems, the mainstream detection technology of electricity theft is based on deep learning, which has the risk of leaking privacy during centralized processing. Meanwhile, due to the insufficient number of samples, the models trained by local data face the problem of poor generalization and usability. Aiming to the above challenges, in this paper, we propose an electricity theft detection method based on asynchronous federated learning that supports personalized privacy protection, enabling the participants to model jointly without data leaving the local devices. Combining the asynchronous federated learning mode and the personalized privacy-preserving mechanism, an asynchronous aggregation algorithm considering both asynchronous staleness and the privacy budget difference is designed, achieving the balance of participants' personalized privacy-preserving requirements and model performance effectively. The experiments prove that the proposed method can achieve personalized privacy protection and ensure the generalization performance of the model at the same time. Meanwhile, the global model can converge rapidly and stably by utilizing our algorithm.