Adversarial-aware Machine Learning for Enhancing Smart Grid Security

Smart grid systems are essential components of modern power infrastructures, where machine learning has found wide-ranging applications, particularly in the development of intrusion detection systems. However, such systems remain vulnerable to adversarial attacks, including data poisoning techniques like label flipping attack. This paper proposes a novel, adversarial-aware machine learning-based intrusion detection system that is robust against label flipping attacks, thereby enhancing the integrity and reliability of the energy distribution network. The proposed solution was evaluated using the ICS cyber-attack dataset from the University of Queensland, with multiple models undergoing progressive testing. Initially, the base model employed three key algorithms: Random Forest, K-Nearest Neighbors, and XGBoost. These models were subjected to simulated label flipping attacks with varying intensities (10%-50%), which caused a significant decline in performance. To counter this, a countermeasure technique was integrated into the system, successfully restoring model accuracy and achieving a consistent accuracy of 75.96% across all attack intensities. Among the algorithms, Random Forest demonstrated the greatest resilience, showing a notable recovery after mitigation. This study underscores the critical role of preventive defense strategies and robust machine learning algorithms in safeguarding smart grids against cyber threats.