Deep learning image adversarial sample defense method and device, equipment and storage medium

Abstract

The invention provides a deep learning image confrontation sample defense method, device and equipment and a readable storage medium. The method comprises the following steps of acquiring a deep learning model and an image training data set with a plurality of original images; randomly dividing the image training data set into a plurality of batches to obtain a plurality of small-batch samples, and performing random image transformation processing on part of the small-batch samples according to a random probability; training the deep learning model by using each small-batch sample until all the small-batch samples complete the training of the deep learning model; and repeating the training till the deep learning model completes the training. According to the defense method provided by the invention, rich difference changes can be brought to the input data through fusion and mixing of homogeneous images in different modes, so the confrontation space of the data set found by the attack method is difficult to align with the changes, equivalently, the original confrontation space is disturbed, and thus a relatively good defense effect is generated.

Description

technical field [0001] The present invention relates to the technical field of artificial intelligence, in particular to a deep learning image defense method, device, equipment and storage medium. Background technique [0002] With the rapid advancement of computer technology, deep learning has developed explosively. Especially in the field of computer vision, deep learning has been widely studied and has been fully applied in many fields such as pedestrian re-identification and traffic monitoring. However, deep learning models also face severe security issues, especially vulnerable to adversarial examples. Adversarial examples are a type of malicious images formed by artificially adding slight perturbations to normal images, which can mislead image classification models based on neural networks, causing the neural network to completely change its classification of images, and then be exploited by attackers. This will have a great negative impact on the reliability and secu...

AI Technical Summary

Problems solved by technology

However, deep learning models also face severe security issues, especially vulnerable to adversarial examples

Adversarial examples are a type of malicious images formed by artificially adding slight perturbations to normal images, which can mislead image classification models based on neural networks, causing the neural network to completely change its classification of images, and then be exploited by attackers. This will have a great negative impact on the reliability and security of the system

These adversarial examples are only slightly perturbed, so that the human visual system cannot detect this perturbation

Especially in the field of person re-identification, it is quite important to check its adversarial defense capabilities, because the insecurity of person re-identification models may cause serious losses, for example, criminals may use adversarial jamming to fool surveillance systems

Images