The smart grid integrates Information and Communication Technologies (ICT) into the traditional power grid to manage the generation, distribution, and consumption of electrical energy. Despite its many advantages, it faces significant challenges, such as detecting abnormal behaviours in the grid. Identifying anomalous behaviours helps to discover unusual user power consumption, faulty infrastructure, power outages, equipment failures, energy thefts, or cyberattacks. Machine learning (ML)-based techniques on smart meter data has shown remarkable results in anomaly detection. However, traditional ML-based anomaly detection requires smart meters to share local data with a central server, which raises concerns regarding data security and user privacy. Server-based model training faces additional challenges, such as the requirement of centralised computing power, reliable network communication, large bandwidth capacity, and latency issues, all of which affect the real-time anomaly detection performance. Motivated by these concerns, we propose a Federated Learning (FL)-based smart grid anomaly detection scheme where ML models are trained locally in smart meters without sharing data with a central server, thus ensuring user privacy. In the proposed approach, a global model is downloaded from the server to smart meters for ondevice training. After local training, local model parameters are sent to the server to improve the global model. We secure the model parameter updates from adversaries using the SSL/TLS protocol. Using standard datasets, we investigate the anomaly detection performance of federated learning and observe that FL models achieve anomaly detection performance comparable to centralised ML models while ensuring user privacy. Further, our study shows that the proposed FL-based models perform efficiently in terms of memory, CPU usage, bandwidth and power consumption at edge devices and are suitable for implementation in resourceconstrained environments, such as smart meters, for anomaly detection.
