Electricity Theft Detection Based on Federated Learning

Electricity theft not only seriously endangers the stable transmission of the electricity grid, but also disrupts the normal order of electricity consumption due to data errors that bring great disturbance to the judgement of the grid. Most machine learning algorithms for detecting electricity theft are trained by concentrating data. However, in practice, concentrating so much data for training is bound to have data privacy concerns and troubles. Therefore, this paper proposes the use of federated learning for electricity theft detection, where federated learning ensures that the power company does not need to upload data, but only needs to achieve global iteration of the model by uploading the model parameters. In this paper, FedAvg algorithm and FedProx algorithm are used to implement the task of electricity theft detection for wide and deep models respectively. On this basis, in order to better protect user privacy, this paper encrypts the aggregation process of federated learning using differential privacy and implements the FedAvg and FedProx algorithms based on differential privacy, respectively.