Adversarial Defense Based on Mimic Defense and Reinforcement Learning for Power Vision Task in Smart Grid

The utilization of deep learning-based vision technology has emerged as a crucial approach in the digitalization of new power system. Nevertheless, the extensive implementation of deep learning has progressively revealed its deficiencies in terms of interpretability and robustness, resulting in increased security incidents. Notably, deep learning models possess inherent vulnerabilities and are prone to adversarial attacks, as the system may produce inaccurate judgments when introducing disturbances to the input data. This situation imposes substantial risks on security-oriented new power system. Hence, to fortify the adversarial defense capability of vision technology, this research establishes a power vision security system that leverages reinforcement learning and mimic defense. The system incorporating seven core modules facilitates automatic model selection based on value and credit, determination of visual analysis results through consistency, and model updates derived from hard samples. Consequently, the adversarial defense capability of the power vision security system is substantially enhanced from an algorithmic standpoint, facilitating the perception and comprehension of unknown threats. To demonstrate the effectiveness of the proposed power vision security system, this paper employs insulator detection as a case study and conducts simulated attack experiments utilizing adversarial samples.