A method, system and terminal for generating highly transferable adversarial samples

Abstract

The invention discloses a method, system and terminal for generating highly transferable adversarial samples, which belong to the field of deep learning technology. The low-frequency information of the first target image is added to the randomly selected original image to obtain the first sample; the first sample is obtained randomly from the model pool. Select a first-level sampling model; input the first sample into the first-level sampling model; iteratively calculate the aggregated gradient of the sampling model based on the low-frequency information of the first target image added to the original image, and normalize the aggregated gradient; update the sampling model based on the sign of the gradient The generated adversarial samples are constrained on the value of the adversarial samples, and the first-level temporary adversarial samples are output as the input of the lower-level sampling model; the above steps are repeated to obtain the final generated samples. This application has a multi-level sampling model, which improves the transferability of the final generated adversarial samples. At the same time, the samples of the input sampling model are superimposed with target low-frequency information different from its own category, which can further improve the transferability of adversarial samples.

Description

technical field [0001] The present invention relates to the technical field of deep learning, in particular to a method, system and terminal for generating highly transferable adversarial samples. Background technique [0002] In recent years, with the rapid development of deep learning, its intelligent applications in image recognition, face recognition, unmanned driving, target detection, medical treatment and other fields have become a reality. Compared with traditional methods, deep learning has shown extremely high performance in various fields. However, some studies have shown that deep learning is vulnerable to adversarial attacks due to the existence of adversarial samples. In today's highly informatized and intelligent world, the security guarantee of deep learning algorithms is an unprecedented difficulty and pain point. [0003] The adversarial samples generated by the attacker through carefully designed perturbations are indistinguishable from the real samples t...

AI Technical Summary

Problems solved by technology

[0008] The purpose of the present invention is to overcome the problem of low transferability of adversarial samples generated in the prior art in model attacks, and provide a method, system and terminal for generating adversarial samples with high transferability

Images