A low-cost adversarial network attack sample generation method based on gan

Abstract

The invention discloses a method for generating low-cost adversarial network attack samples based on GAN, comprising: step 1, encoding the attributes and labels of the samples in the sample set; step 2, calculating information gain according to the attributes of the encoded samples; 3. Adversarial example generation model training; step 4, generating adversarial examples for attack. Based on the basic idea of ​​GAN, the present invention can realize the automatic selection of disturbance features by calculating the attributes, labels and information gain training models of samples, and generate efficient adversarial network attack samples at the minimum cost.

Description

technical field [0001] The invention relates to the field of network security, in particular to a GAN-based low-cost adversarial network attack sample generation method. Background technique [0002] In recent years, artificial intelligence algorithms have been gradually applied in the field of network security, and have shown good performance in malware detection, intrusion detection, and vulnerability mining. However, because artificial intelligence algorithms are vulnerable to adversarial attacks, systems such as malware detection and intrusion detection are vulnerable to threats from malicious attackers. To this end, researchers have proposed methods to defend against adversarial attacks, among which methods based on generative adversarial network (GAN) are the majority, and have achieved good defense effects. However, most of these methods have the following three problems: [0003] 1. Most methods focus on how to use GAN to improve the accuracy of the detection algor...

AI Technical Summary

Problems solved by technology

[0003] 1. Most methods focus on how to use GAN to improve the accuracy of the detection algorithm, and do not pay attention to the ability of GAN-based adversarial attacks to network attack detection algorithms;

[0004] 2. Most of the methods are derived from the generation methods of adversarial samples for images. These methods do not take into account the differences between image samples and network data samples, that is, in network data, different attributes have different importance to classification results;

[0005] Third, these methods do not take into account the cost of generating adversarial samples

Images