A novel strategy for locational detection of false data injection attack

Abstract

State estimation algorithms furnish an effective approach in monitoring and control of critical infrastructures like smart grid in real-time. Recently, false data injection attack (FDIA) has shown its detrimental effects by bypassing the conventional bad data detection test. The current trend in the detection technique of such class of attack remains localised in their presence detection within the measurement set whereas their exact locations of intrusions remain unknown. To palliate this issue, this work showcases an effective implementation of deep learning models to determine the exact intrusion points in real-time. Such novel deep learning models along with the conventional bad data detector can effectively identify the presence of any unstructured and structured false data injections within the measurement set, hence providing a cost-effective strategy. These deep learning models can effectively capture the inconsistency with co-occurrence dependency of the potential attack vectors within the measurement set, thus providing a multilabel classification approach. Moreover, such class of deep learning architectures being model free endow the grid operators with real-time detection without any preliminary statistical assumptions of the grid. An extensive analysis on the standard IEEE test bench showcases the efficacy of such class of detection policy undergoing varying attack and noise margins.