Intelligent confrontation sample generation method and system based on optimization algorithm and invariance

Abstract

The invention belongs to the technical field of image recognition data processing, and particularly relates to an intelligent adversarial sample generation method and system based on an optimization algorithm and invariance. The method comprises the steps: collecting original image data with a correct label; constructing a neural network model for adversarial sample generation and a model loss function, and optimizing adversarial disturbance between an original input image and a corresponding output adversarial sample by maximizing the model loss function; based on original image data and a neural network model, an Adazief iterative quick gradient method and a cutting invariance method are used for iterative solution, and a finally generated adversarial sample is obtained according to an iteration termination condition. From the perspective that the generation process of the adversarial sample is similar to the neural network training process, the convergence process is optimized through the Adazief iteration quick gradient method, the over-fitting phenomenon in the adversarial attack is avoided by using the cutting invariance, the adversarial sample with better mobility can be generated, the robustness of the network model is improved, and the practical scene application is facilitated.

Description

technical field [0001] The invention belongs to the technical field of image recognition data processing, and in particular relates to a method and system for generating intelligent adversarial samples based on optimization algorithms and invariance. Background technique [0002] In the field of image recognition, experimental results on relevant standard datasets show that the recognition ability of deep neural networks can reach or even exceed the human level. However, researchers have found that deep neural networks are fragile. For example, Szegedy et al. first discovered an interesting property of deep neural networks: adding small perturbations imperceptible to humans to the original clean image can make deep neural networks give wrong outputs with high confidence. The perturbed image is an adversarial example; although the existence of adversarial examples seriously affects the safe use of deep neural networks, adversarial examples with strong attack performance can ...

AI Technical Summary

Problems solved by technology

However, these existing methods often show a low attack success rate in a black-box environment, especially for adversarially trained networks

Images