Improving FDI Detection for PMU State Estimation Using Adversarial Interventions and Deep Auto-Encoder

Concerns have been voiced about the growing significance of cyber-threats, especially in light of the potentially dire repercussions of false data injection (FDI) assaults. This work investigates FDI detection in phasor measurement units (PMU), focusing on instances where an attack can be launched simply by compromising one unit. Simulated post-processing adversarial interventions i.e., noise and non-linearity were introduced to train and fortify the system against possible attacks and to render it resilient to perturbations. By learning complex non-linear patterns from the data, a deep de-noising auto-encoder model is used to de-noise and learn genuine feature representations, improving overall reliability. The suggested framework performs better than conventional machine learning and 1-D CNN models when it comes to precisely estimating intrusion, as shown by a comparison study. By using an integrated strategy, power system monitoring and control become more accurate and resilient, successfully tackling the changing issues faced by contemporary electrical grids. The proposed adversarially robust framework is evaluated using Monte-Carlos simulations and on varying load conditions to better comprehend the impact of adversarial interventions on the FDI detection accuracy under different load characteristics and attack scenarios. The proposed framework yielded an average 98.3% in Monte Carlo simulations and an average of 96.5% accuracy under varying load conditions. Surpassing the conventional ML and 1-D CNN algorithms in successfully identifying FDI attacks under adversarial vulnerability.