Research requirements of robustness evaluation technology of artificial intelligence model in power visual scene

For trusted AI, the AI academic community has set up related technical communities, including AcmFacct, Fair & Responsible AI CHI2020, RAI ICLR, etc. However, the sample set adopted by the research in the field of general artificial intelligence technology lacks the coverage of power vision data, which is divorced from the actual power vision scene, and ignores the influence of the evil, diversity and complexity of the outdoor environment of power vision on the model decision credibility, which makes the existing credibility evaluation tools unable to be directly applied to the actual power vision scene. At the same time, there is a lack of critical technical research and support within the industry to accurately measure the model's ability to provide reliable service in real power business scenarios. Through the effective evaluation of the adversarial robustness of the power vision model under three different strategies: white box, grey box and black box, the performance of the power vision model under adversarial attack is verified, and the security and reliability of the power vision model in key power scenario applications are guaranteed, the security risks of the model are reduced, and the reliability of model users on model output decisions is improved. In turn, it can enhance the confidence of users, accumulate more high-quality data in the process of frequent use and sustainable training, and improve the ability of artificial intelligence models, which complement each other, and finally make artificial intelligence technology fully land in the power grid, so that the power grid can enter the intelligent era as soon as possible.