An Advanced Defense Mechanism for Detecting False Data Injection Attacks in Cyber-Physical Systems through Feature Selection Based-Machine Learning Algorithms

False Data Injection Attacks (FDIAs) are among the most common cyber threats targeting measurement data and sensor readings in power systems. These attacks pose a significant risk to the integrity of power system data, as they can evade conventional Bad Data Detection (BDD) mechanisms. This evasion can lead to incorrect decisions by power system operators and control centers during Power System State Estimation (PSSE), potentially resulting in serious damage to the system topology and economic losses. While previous approaches have primarily focused on enhancing BDD algorithms, there remains a significant gap in effectively detecting stealthy FDIA attacks. This paper presents an approach that leverages Machine Learning (ML) algorithms, including supervised models such as Decision Trees (DT) and K-Nearest Neighbours (KNN), alongside ensemble classifiers like Random Forest (RF), XGBoost (XGB), and AdaBoost (ADABO). A feature selection technique is applied post-training of the XGBoost model to identify the most relevant features, thereby improving the performance of the classifiers. The proposed method has been rigorously tested on the IEEE-14 bus system, demonstrating significant improvements in detection accuracy and computational efficiency. The results are discussed in detail across various metrics and time-based evaluations, highlighting the robustness of our approach.