LAUSR

Abstract With the influx of more advanced and more connected computing and control devices, the electric power grid has continuously evolved to rely on communication networks for efficient operation and control. A challenge with these new technologies is that they may introduce new and unforeseen avenues of access, making the grid more susceptible to cyber attacks. False Data Injection Attacks (FDIA) are a particular type of attack that aims to cause disruptions in the operation of the power grid by affecting the feedback mechanism to control the grid. This is carried out by modifying the measurements which enable a state estimator to approximate the state of the system. These attacks are designed in such a way that they preserve the system equations on which the state estimator operates; therefore, they cannot be detected by a simple residual-based detection mechanism. In this paper, we propose monotonic attention based auto-encoders, an unsupervised learning technique to detect FDIAs. The auto-encoder is trained under normal operating conditions, and we hypothesize that it will produce outputs which are close to the true system values at normal operation even if the measurements are modified by an adversary. Based on this hypothesis, that high reconstruction error occurs for the attacked conditions, the intrusion detection is performed by a threshold mechanism using Precision-Recall curve. We validate the efficacy of our proposed attention-based auto-encoder anomaly detector (A3D) over other variants of auto-encoders such as ANN and RNN based auto-encoders, and a few supervised learning techniques, by performing FDIAs on a IEEE 14 bus system.