Supervised Learning Models empowered by XGBfs and PCA for Unobservable Cyber Attacks detection in Cyber-Physical Systems

Unobservable cyber-attacks, which manipulate measurement data and sensor readings in modern power systems, are commonly known as False Data Injection Attacks (FDIAs). Their ability to bypass traditional Bad Data Detection (BDD) mechanisms poses a significant threat to the integrity of power system informationSuch evasion can cause significant disruptions to system topology and lead to financial losses by misleading power system operators and control centers into making erroneous decisions during Power System State Estimation (PSSE). Detecting these attacks is critical. Therefore, this study proposes an approach to enhance the detection of such stealthy attacks. This study introduces a range of machine learning (ML) algorithms, including ensemble classifiers such as AdaBoost (ADABO) and Random Forest (RF), as well as simpler supervised classifiers like Decision Trees (DT) and K-Nearest Neighbors (KNN). Additionally, this study explores techniques for reducing data complexity and enhancing model performance by selecting the most relevant information from datasets through feature engineering. Extreme Gradient Boosting-based feature selection (XGBfs) is employed to identify the optimal feature subset, while Principal Component Analysis (PCA) is introduced as a second method for feature extraction. These techniques significantly improve model performance. The IEEE-14 bus test system has been utilized to evaluate the effectiveness of the proposed approach, with results analyzed using performance metrics and training durations.