Odel robustness evaluation algorithm based on adversarial perturbations

Deep Neural Networks (DNNs) have performed well in various tasks, but their robustness issues have gradually become prominent with their widespread application. Robustness evaluation is crucial to ensure that the model can maintain performance and stability under various input changes. However, existing robustness evaluation methods are usually difficult to fully capture the performance of the model in complex and dynamic real environments, resulting in a deviation between the evaluation results and the robustness requirements in practical applications. This paper proposes a model robustness evaluation algorithm based on adversarial perturbations, which is a method based on effectively approximating the decision space of the target classifier to find the minimum perturbation required to deceive the model. Based on an in-depth analysis of adversarial sample generation methods, this study uses the perturbation vector generated by unitization as the perturbation direction, gradually adds perturbations to the image along this direction until the model misclassifies, obtains the minimum perturbation and calculates the perturbation as the minimum distance between the image and the decision boundary, so as to more accurately evaluate the robustness of the model, thereby optimizing the problem of inaccuracy of current evaluation methods. Experimental results show that this method performs well in calculating smaller perturbations, improving the accuracy and reliability of model robustness evaluation.