Image classification neural network attack method based on Bezier curve

Abstract

The invention belongs to the field of artificial intelligence and deep learning, and discloses an image classification neural network attack method based on a Bezier curve. The method includes: firstly, building a plurality of image classification neural networks, and loading pre-training parameters; then, selecting one thousand of image samples meeting the same distribution from a classic Image Net data set, and performing parameter fine adjustment on the built neural network by utilizing the image samples to improve the accuracy; randomly generating a Bezier curve on the basis of the original image; searching an optimal curve from a large number of randomly generated Bezier curves by using a differential evolution algorithm; and finally, obtaining an adversarial image sample containing the optimal Bezier curve, and misleading an image classification neural network to classify the adversarial image sample by mistake.

Description

technical field [0001] The invention belongs to the field of artificial intelligence and deep learning, and in particular relates to an image classification neural network attack method based on Bezier curves. Background technique [0002] As we all know, the performance of deep learning based on convolutional neural network has surpassed traditional machine learning methods in the field of artificial intelligence, and plays a leading role in the fields of image recognition, natural language processing and speech processing, especially in the field of image classification, the most advanced Models have outperformed humans. [0003] However, Szegedy et al. found that when adversarial perturbations are added to images, the model becomes fragile. At the same time, these perturbations are almost indistinguishable to humans. In the following years, multiple algorithms for adversarial sample generation showed that adversarial images can achieve a very high accuracy rate when att...

AI Technical Summary

Problems solved by technology

[0009] In order to overcome the shortcomings of the above methods, the present invention proposes an image classification neural network attack method based on Bezier curves, which can solve the problem that the modified image has a great influence on the human recognition effect, and can effectively attack the image classification neural network. network attack

Images