Spatio-temporal graph attention network-based detection of FDIA from smart meter data at geographically hierarchical levels

The power consumption data from residential households collected by smart meters exhibit a diverse pattern temporally and among themselves. It is challenging to distinguish between regular consumer behavior and injected falsified measurements into the data stream with the intent of energy theft or compromising the security of the associated measurement infrastructure. This work identifies the challenges of detecting falsified measurements in smart meter data aggregated at geographically hierarchical levels and proposes a novel graph attention network (GAT)-based unsupervised learning framework to detect false data injection attacks (FDIA) from the moving statistics of the power consumption data in real-time, namely MOVSTAT-GAT. The proposed technique is capable of detecting falsified measurements at both 9-digit and 5-digit ZIP code labels in an unsupervised manner, solely from smart meter power consumption data with no additional meters. Moreover, the proposed technique offers a visualization technique to assist the operator in identifying the localization characteristics of the attack and proposes an automated localization strategy for localized FDIAs. Experiments suggest the effectiveness of the proposed framework, especially for localized FDIA or external anomalies, such as power outages and denial-of-service (DoS). Additionally, a detailed discussion regarding the implementation of MOVSTAT-GAT in the industrial environment has been provided.