Transferable image adversarial sample generation and deep neural network test method and system

Abstract

The invention provides a transferable image adversarial sample generation and deep neural network test method and system, and aims to test the robustness of a deep neural network model in a real environment. Firstly, data are collected, and a proxy model with the same function as a to-be-tested model is obtained through training. Diversified neighbor samples of the input image sample are then generated. And obtaining the gradient of the loss function relative to the neighbor image sample on the agent model, carrying out standardization operation on the gradient, and taking the average of the gradient as the disturbance direction. And finally, pixel values which do not conform to constraint conditions are cut according to dynamic constraints. And repeating the steps to reach the maximum number of iterations to obtain transferable image confrontation samples, and testing the to-be-tested model by using the transferable image confrontation samples. According to the method, the sample generation method is innovated through the diversification, gradient standardization and dynamic constraint generation of neighbor image samples, and the quality of the generated image is improved while the generation success rate of the transferable image confrontation sample is improved.

Description

technical field [0001] The invention relates to a test method and system for generating a transferable image confrontation sample and a deep neural network model, and belongs to the field of artificial intelligence testing. Background technique [0002] Although deep neural networks have achieved good results in many computer tasks and practical applications, recent studies have shown that DNNs for image processing tasks are particularly vulnerable to image adversarial examples that add tiny perturbations. These disturbances cannot be detected by humans, but they can cause serious errors in the AI ​​system, which brings security risks to the actual deployment of DNNs. [0003] At present, the security of DNNs mainly tests whether the deep neural network model can correctly predict on the generated image samples by generating image adversarial samples of the model to be tested. Although image adversarial samples exist widely, the threat of image adversarial samples generated...

AI Technical Summary

Problems solved by technology

[0005] (1) The test method for the neural network of transferable image confrontation samples still has a weak ability to find errors, and the instability of the optimization direction is one of the main reasons for the weak ability to find errors;

[0006] (2) The current test method blindly pursues the ability to find errors but ignores the "imperceptibility" constraint of the disturbance, resulting in poor quality of the generated image samples

Images