A Universal Defense Strategy for Data-Driven Power System Stability Assessment Models Under Adversarial Examples

Based on the machine learning (ML) technique, the data-driven power system stability assessment (SA) has received significant research interests in recent years. However, even with a high SA accuracy performance, the data-driven SA models may be vulnerable to adversarial examples (caused by some physical noises or adversarial attacks), which are very close to the original input but can result in a wrong SA result. To solve such threat, this article first proposes a universal defense strategy for the ML-based SA models based on the randomized smoothing algorithm to resist the adversarial attacks. Second, this article proposes an effectiveness index for the proposed universal strategy to quantify the maximum ability of resistance to adversarial examples. Moreover, this article provides the tight mathematical proof for the effectiveness index under the hard smoothing, soft smoothing, and binary scenarios. Simulation results verify that the proposed defense strategy can effectively resist the adversarial examples and the proposed effectiveness index can provide a formal robustness guarantee for real-time power system SA applications.