Adversarial Defence Model With Improved U-NET and Generative Adversarial Network
MO Ruilin
Abstract：Aiming at the problems of weak generalization,image denoising and feature extraction of existing countermeasure defense methods,an attack defense model DAU-NET-GAN is proposed,which integrates improved U-NET and generated counter-measure network.By introducing channel attention mechanism and non-local mean filter in U-NET downsampling process,the in-fluence of small perturbations on the image can be reduced as much as possible,the ability of the model is improved to extract the key features of the image,and soft attention mechanism is added to the jump connection part of the generation network,so as to avoid feature redundancy and realize the reconstruction of the adversus-sample.The GCE loss function is used to replace the cross entropy loss function in the training process,which improves the robustness of the model to the opposing samples.The experimental results show that the proposed model can effectively defend against the counter samples generated by various attacks,and the de-fense success rate on MNIST and CIFAR10 data sets can reach 98.96%and 83.88%,which has good general defense effect.
Keywords：deep neural networkgenerative adversarial networksadversarial attackadversarial defence
Publication Date：2025-11-20
Online Publishing Date：2026-01-28(First online date of this platform, not the publication date of the document)
Pages：6( 3168-3173 )