Security Risk Assessment of Image Classification Model Based on ANP-TOPSIS

Recently, machine learning algorithms have been widely used in the fields of image processing, network security and natural language processing, etc., profoundly affecting human life. However, machine learning algorithms have the characteristics of uncertain output, vulnerability to adversarial attacks, and unexplained decision-making processes, which seriously threaten the security of machine learning-based face recognition, Malware detection, and autonomous driving. Hence, it is imperative for the security practitioners to evaluate algorithm security to ensure that security needs are met. In this article, the authors propose a set of security assessment index systems and methods for machine learning algorithms for image classification scenarios: Refer to the security specification of machine learning algorithms and requirements to construct the security index system of image classification model. Furthermore, The Analytic Network Process(ANP) is applied to quantify the index weights and the Technique for Order Preference by Similarity to an Ideal Solution(TOPSIS) is applied to screen the optimal model, and finally the sensitivity analysis is applied to prove the stability of the proposed method. Experimental results show that this method has certain value and effect in assessing the security and model screening of image classification models.