Method for enhancing image classification robustness

Abstract

The invention belongs to the field of computer software, and particularly relates to a method for enhancing image classification robustness, which can enhance the anti-interference performance and robustness of a classification model. The method aims at defending attacks of most traditional white-box confrontation samples. The method mainly comprises an adversarial sample detection network generation module which is used for constructing an adversarial sample detection network by adding a neural network layer on the basis of an original classifier, wherein the network mainly recognizes adversarial samples; a judgment threshold generation module which is used for finding a proper judgment threshold of the adversarial sample detection network by using a common adversarial sample method; and an enhancement model generation module which is used for further training to obtain an enhanced image classifier by combining a classification result of the detection network on the basis of the classification of the image classifier of an original model, and finally carrying out image classification by utilizing the enhanced image classifier so as to improve the robustness of the classifier.

Description

technical field [0001] The invention belongs to the field of computer software, in particular to an image classification model robustness enhancement evaluation method, which can enhance the anti-interference and robustness of the classification model. Background technique [0002] In recent years, deep learning has been widely used in China, and has achieved good results in image classification, face recognition, and language processing. Especially in image recognition, it can even match the performance of human beings, and the existing technology can reach a recognition rate of more than 99%. However, most researchers care more about the performance of the model (such as the correct rate), but ignore the fragility and robustness of the model. Foreign Szegedy et al. found in experiments that by adding disturbances that are difficult to distinguish with the naked eye to the image, the final model cannot obtain correct classification results. Then Szegedy et al. proposed to...

AI Technical Summary

Problems solved by technology

These algorithms can generate good perturbations, which can cause the model to classify incorrectly, or classify into the classification desired by the attacker. These problems have caused people to pay attention to whether personal safety can be guaranteed when deep learning is applied to real-world scenarios.

Images