A deep learning classification method with the function of defending against adversarial sample attacks

Abstract

The invention discloses a deep learning classification method with the function of defending against adversarial sample attacks. The device for realizing the classification method includes: an attack generation model AG, an adversarial sample discrimination model D, and a classification model DNN. The specific methods are as follows: (1) Normal use The data set trains DNN, and the classification accuracy is greater than the preset value, and the training of DNN is stopped; (2) Alternately train the parameters of AG and D until AG‑D achieves Nash equilibrium; (3) Alternately train the parameters of AG and DNN until AG-DNN realizes Nash equilibrium; (4) judge whether the adversarial example discrimination model D and the classification model DNN reach Pareto optimality, if yes, DNN training is completed, and step (5) is performed; otherwise, return to step (2); ( 5) Input the sample to be classified into the trained classification model DNN to obtain the classification result. The invention can effectively solve the vulnerability of the classification model in the face of confrontation samples in practical classification applications, and improve the robustness of the model performance.

Description

technical field [0001] The invention belongs to the research field of deep learning algorithms and models in the field of artificial intelligence, and specifically relates to a deep learning classification method with the function of defending against adversarial sample attacks. Background technique [0002] At present, deep learning has attracted the attention of a large number of scholars and research institutions. Through the continuous improvement of hardware performance, deep learning has been widely used in object detection, image semantic segmentation and understanding, knowledge graph, data generation and other fields. Deep learning is one of the most commonly used technologies for artificial intelligence machine learning, and adversarial attacks on deep neural networks are a serious security risk. Adversarial attack is defined as a malicious attack in which an attacker obtains an adversarial sample by adding carefully designed small perturbations to the original dat...

AI Technical Summary

Problems solved by technology

In addition, adversarial examples are a solution to nonlinear and non-convex optimization problems of deep neural network models, and it is necessary to have a good method to describe these complex optimization problems; from the perspective of defense, most defense research focuses on the defense of a single attack, through Capture the attacker's adversarial samples and start the adversarial learning of the deep neural network to defend against the attack. Therefore, adversarial training that relies on a small number of existing adversarial samples is not universal

Images