A Voting-Based Machine Learning Strategy to Detect False Data Injection Attack in Cyber-Physical Power Systems

Integrating the information and communications technology (ICT) infrastructures with the physical layer of the electrical grid puts the power system at the risk of a wide range of cyber intrusions. Cyber-physical power systems (CPPS) require effective and applicable prevention, detection, and mitigation strategies to defend the system against malicious attacks. In this process, precise attack detection is one of the vital phases. This paper designs a novel voting-based detection tool to identify cyber intrusions in the system. In the considered cyber-attack, the attacker attempts to inject a wide range of false data into the phasor measurement units (PMU) to simulate false short circuit conditions in the system. The proposed detection strategy utilizes several machine learning (ML) algorithms like ensemble learning, discriminant analysis, naive bayes, feedforward neural network (FNN), recurrent neural network (RNN), k-nearest neighbors (KNN) classification, support vector machine (SVM), and decision tree. The voting-based technique calculates the average output based on the performance of the detectors and can recognize FDI attacks from real short circuit faults. Training features are selected optimally among several mechanical and electrical features of the system with maximum relevancy and minimum redundancy (MRMR) method. The simulations on the IEEE 39 bus test system illustrate the performance of the proposed detection tool for each type of short circuit fault.