Defense Methods for Adversarial Attacks Against Power CPS Data-Driven Algorithms

The integration of large-scale power electronic devices has introduced a large number of strong nonlinear measurement/control nodes into the system, gradually transforming the traditional power system into a cyber physical system (CPS). Many system problems that were originally solved by model-driven methods have had to be analyzed using data-driven algorithms due to limitations such as dimensional disasters. However, the inherent flaws of data-driven algorithms introduce new risks to the safe and stable operation of the system, which attackers can exploit to launch adversarial attacks that may cause system power outages and even instability. In response to the potential adversarial attacks on data-driven algorithms in power CPS, this paper proposes corresponding defense methods from such three aspects as abnormal data filtering and recovery, algorithm vulnerability mining and optimization, and algorithm self interpretability improvement: abnormal data filter, GAN-based vulnerability mining and optimization method, data knowledge fusion model and its training method. The effectiveness of the proposed method is verified through case analysis.