Privacy-Aware Split Learning Based Energy Theft Detection for Smart Grids

ورقة مؤتمر

2022

تم النشر فى: https://link.springer.com/chapter/10.1007/978-3-031-15777-6_16

Energy thefts are one of the critical attacks that often cause high revenue losses for utility companies around the world. Effective detection of such attacks is very important and must be implemented to comply laws and regulations that govern users’ privacy. Current detection approaches rely on significant amounts of raw fine-grained smart meter data and generally do not consider privacy. On the other hand, most privacy-preserving machine learning (PPML) approaches, such as homomorphic ML and federated learning, are not well suited to the smart grid environment due to their processing complexity and communication overheads. Therefore, our contributions in this work are twofold: first, we propose an enhanced privacy-preserving detection model for energy thefts using the concept of Split Learning. Subsequently, since the classical Split Learning cannot be directly applied in the smart grid (SG) environment due to its communication overhead, we introduce a new variant of Split Learning that is more communication-efficient and suits the smart grid environment. The proposed model can ensure two advantages over the existing techniques. First, the use of Split Learning enables the training of a detection model without any need for raw data. This helps in achieving data privacy. Second, the splitting of the detection model allows the system to be more robust against honest-but-curious adversaries. Our evaluations show that the proposed detection model can ensure better privacy protection and communication efficiency, which are essential for smart grid, without compromising detection accuracy.