A Federal Learning Framework for Privacy-protected Distributed Power Theft Detection

Abstract

Aiming at the security problem of electricity consumption data, a federal power theft detection method is proposed. The method does not need to upload the power consumption data to the data center, and can extract the power consumption features through local training, which reduces the risk of data leakage. Compared with other federal detection methods, this method uses the user's local data to train the model, which improves the privacy protection ability and anomaly detection accuracy. In the comparison experiment with the baseline method, the F value of the proposed method reached 94%, and the false positive rate and false negative rate were lower than the existing federal anomaly detection method.