RE4ETD: A Relative Entropy Optimization-Based Method for Efficient Electricity Theft Detection With Dual-Privacy Preservation

Transmitting residents’ electricity usage data from the Edge Server (ES) to the Detection Server (DS) for Electricity Theft Detection (ETD) poses significant privacy risks to electricity usage data in smart grid. To mitigate this risk, existing studies primarily employ encryption methods to perform encrypted detection on the electricity usage data. However, these approaches assume that the ES is trustworthy while the DS may be malicious, overlooking the possibility that the ES could also steal the ETD model from the DS. Moreover, none of these methods address the data imbalance problem in privacy-preserving scenarios, leading to inefficient ETD performance. Thus, we propose RE4ETD, a method to protect the privacy of electricity consumption data and ETD model simultaneously (dual-privacy-preserving) with effective performance. RE4ETD consists of three modules: data transformation, feature extraction, and post-detection. First, on the Edge Server (ES) side, we employ relative entropy optimization to project private electricity consumption data into a distribution $ Q $ that is significantly different from the original private distribution. Then, randomly sampling from $ Q $ and sent these samples to DS. Subsequently, the DS extracts features and transmits these features back to the ES. Finally, the ES uses these features to obtain the detection results. We employ split learning and customized convolutional kernels to train these interconnected modules. Additionally, we utilize ensemble learning to mitigate the persistent data imbalance issue in the ETD scenario. Extensive experiments demonstrate that, compared to previous methods, RE4ETD achieves average improvements of 5-8 percentage points in F1-score and AUC, respectively, with low communication costs.