Robustness evaluation and enhancement system of artificial intelligence image classification model

Abstract

A robustness evaluation and enhancement system of an artificial intelligence image classification model comprises a white box evaluation module, a black box evaluation module and a defense enhancementmodule, the white box evaluation module obtains a to-be-evaluated model and selected evaluation indexes from a user, the attack resistance of the model is evaluated from all aspects according to multiple different indexes, and the defense enhancement module is used for enhancing the robustness of the model and calculating the scores of all indexes and the total score of robustness; and the blackbox evaluation module obtains an output result of the to-be-evaluated model from the user and compares the output result with the correct label to obtain an evaluation result. A plurality of black boxevaluation means is provided, and the robustness of the model is evaluated from the perspective of black boxes; the defense enhancement module is internally provided with a plurality of robustness improvement means. A to-be-enhanced model and selected defense enhancement method information are acquired from the user and robustness enhancement is performed on the model uploaded by the user by using a corresponding defense enhancement method. The robustness evaluation process of the whole model is optimized through multiple robustness evaluation indexes, so that the model is defended through multiple built-in technologies while different methods can be compared and evaluated more conveniently, accurately and comprehensively, and the robustness of the model is improved.

Description

technical field [0001] The invention relates to the technology in the field of artificial intelligence security, in particular to a system for evaluating and enhancing the robustness of an artificial intelligence image classification model. Background technique [0002] At this stage, the image classification model based on deep learning can achieve high accuracy, but recent studies have shown that by adding artificially constructed small perturbations to normal samples, there is a high probability that the model will misjudge, and such samples are called adversarial examples. Adversarial examples and their mobility make it particularly important to ensure the robustness of the model in an adversarial environment. However, at this stage, there is still no standard evaluation method for the robustness of the model. Different researches on the robustness of the model often use completely different evaluation indicators, so they cannot be compared with each other, thus hinderi...

AI Technical Summary

Problems solved by technology

[0003] Aiming at the problems of non-standard evaluation methods, various indicators, and inability to compare and judge between different evaluations existing in the robustness evaluation of the image classification model at the present stage, the present invention proposes a robustness evaluation and enhancement of an artificial intelligence image classification

Images