A method for generating adversarial attack samples based on generative adversarial networks

Abstract

The present invention proposes a method for generating an adversarial attack sample based on a generative adversarial network, including: S1. collecting training sample images; S2. preprocessing and classification definition; S3. constructing a training framework, which includes a generator, a discriminator and a target Attack the network; S4. Generate random noise and random condition vector based on the training sample image after preprocessing and classification definition; S5. Use random noise and random condition vector as the input of the generator in the training framework, and the generator of the training framework Train with the discriminator, and save the training frame parameters after training; S6. The training is completed, and an adversarial attack sample is generated. The method proposed by the invention has strong data learning ability and high robustness, overcomes the drawbacks of existing methods based on optimization equations and matrix metrics in simple pixel space, and lacks learning ability for data distribution, and improves adversarial attack samples production quality and efficiency.

Description

technical field [0001] The present invention relates to the technical field of generating adversarial attack samples, and more specifically, to a method for generating adversarial attack samples based on generative adversarial networks. Background technique [0002] Deep learning occupies a core position in today's rapidly developing field of machine learning and artificial intelligence. However, modern visual deep neural networks (DNN) cannot effectively resist attacks from adversarial attack samples. The perturbed input samples are very vulnerable, and this kind of adversarial attack samples can mislead the deep neural network to produce wrong output results. [0003] At present, the academic community has proposed a variety of attack strategies for generating adversarial attack samples. In the research of generating adversarial attack samples, a series of available and effective attack models have been proposed. These models can be classified into the following categories...

AI Technical Summary

Problems solved by technology

[0010] In order to overcome the defects that most of the existing adversarial attack sample generation methods are based on optimization equations and simple pixel space matrix metrics, lack of learning ability for data distribution, and have low robustness, the present invention provides a method based on generative adversarial network The generation method of adversarial attack samples improves the robustness of classifiers and the quality and efficiency of adversarial attack samples generation, and promotes the development of adversarial attack research field

Images