P2Detect: Electricity Theft Detection With Privacy Preservation for Both Data and Model in Smart Grid

Abstract

Electricity Theft Detection (ETD) based on deep learning can detect abnormal electricity consumption behaviors by analyzing user historical data. However, existing ETD schemes conduct neural network inferencing in plaintext without strong privacy guarantees, moreover, none has considered the privacy of proprietary models from electric utilities. To address those issues, we introduce p2Detect, a privacy-preserving ETD scheme via neural network inferencing over encrypted data. p2Detect works in a two-party setting where model is held by Detection Server (DS) and data are held by Edge Server (ES) on behalf of multiple Smart Meters (SM). First, we design a homomorphic-encryption-friendly model and its corresponding training methods based on region search and clustering; Second, our customized cryptographic protocols enable privacy-preserving yet efficient ETD inferencing over encrypted data, without revealing private inputs to DS or revealing model parameters to ES. The performance evaluations demonstrate: (i) Model optimizations achieve AUC as high as 0.763, where AUC is a robust evaluation criterion for inferencing ability of the model on unbalanced dataset. (ii) p2Detect takes 73.3ms to execute ETD per SM on average. (iii) Communication cost between DS and ES is 196.76KB per SM. Therefore, p2Detect preserves strong privacy while maintaining high accuracy and efficiency.