Power grid security analysis machine learning model testing methods, devices, equipment and media

Abstract

The purpose of the present invention is to provide a method, device, equipment and medium for testing a machine learning model for power grid security analysis. The method includes: obtaining a fault sample set of power grid operation, the fault sample set including original samples as input to the machine learning model for power grid security analysis; applying perturbations to the original samples to generate a first adversarial sample; with the goal of minimizing the difference between the first adversarial sample and the original sample, and with the output of the wrong classification result of the machine learning model for power grid security analysis as a constraint, iteratively optimizing the first adversarial sample to obtain a final attack sample; inputting the final attack sample into the machine learning model for power grid security analysis to obtain a first classification result; and comparing the first classification result with the correct result to obtain a test result. By automatically generating adversarial attack samples and inputting them into the model, the robustness and security of the model decision are tested, the effect of the adversarial sample attack is quantitatively evaluated, the model decision risk is discovered, and the security of the model application is improved.

Images