Abstract: The rapid development of artificial intelligence has deeply penetrated various industries,with deep neural networks (DNN) being widely applied in multiple fields.However,the popularity of the application exposes the serious threat of adversarial attacks on the vulnerability of intelligent models.Adversarial attacks can lead to model failure,particularly in critical areas such as commercial and military security,potentially resulting in severe consequences.These attacks work by crafting carefully designed inputs to disrupt the normal functioning of models,thus compromising system security and reliability.To comprehensively and scientifically evaluate the robustness of different model algorithms,a quantitative evaluation framework is proposed based on Principal Component Analysis (PCA) for the first time,which encompasses several key evaluation metrics,including misclassification rate,imperceptibility,and attack efficiency.By testing over 20 model algorithms and using PCA to reduce the dimensionality of high-dimensional data,the main evaluation factors are extracted,simplifying the data structure and ultimately producing an overall score for each algorithm.Experimental results demonstrate that the proposed evaluation method is effective and reliable,providing scientific guidance for research on model robustness.

Key words: Principal component analysis, Model robustness, Adversarial examples, Dimensionality reduction