Privacy-preserving federated learning for detecting false data injection attacks on power system

Abstract

False data injection attacks (FDIA) against power system state estimation have been well studied due to its potential threat to real-time energy management. However, the existing works either focus on the case with a sufficiently large local data sets or the local data can be transmitted to a central node, or the case where the system parameters can be easily acquired through communication. This contradicts with real-world applications where the measurement data is geographically distributed and the system parameters are unknown, and also leads to data leakage issues. To solve the problem, a novel FDIA detection algorithm is proposed based on federated learning, through which a global detection model is generated. In the proposed algorithm, the state owners execute a federated learning algorithm using their own data, which avoids massive data transmission and protects the data privacy. Moreover, to prevent the model parameter from exposure to attackers, artificial noise is added into the model estimations to guarantee differential-privacy. Theoretical results show a trade-off in algorithm accuracy and its privacy preserving level. Simulations on the IEEE 30-bus system validate the effectiveness of the proposed mechanism on FDIA detection and its trade-off in accuracy and privacy preserving.