A Method of Enhancing Adversarial Example Transferability Based on NadaMax Update and Dynamic Regularization
SONG Yafei
QIU Wenbo
WANG Yifei
FENG Cunqian
Abstract：To address the problem of insufficient transferability of adversarial examples and inadequate black-box attack capabilities in deep learning models,this study designs an iterative fast gradient method based on the NadaMax optimizer(NM-FGSM).This method integrates the advantages of Nesterov Accel-erated Gradient and the Adamax optimizer,improving the accuracy of gradient updates through adaptive learning rates and lookahead momentum vectors.Additionally,dynamic regularization is introduced to en-hance the convexity of the problem,optimizing algorithm stability and specificity.The experimental re-sults demonstrate that the NM-FGSM is prior to the existing methods under conditions of various attack strategies,particularly in advanced defense scenarios,attack success rate increases by 4％～8％.The dy-namically regularized loss function enhances the cross-model transferability of adversarial examples,there-by further improving black-box attack effectiveness.Finally,points out the way forward for the NM-FGSM algorithm and defense measures,providing a new insight into the security research of deep learning models.
Keywords：transferabilityblack-box attackNadaMax optimizermomentumadaptive learning ratedy-namic regularization
Publication Date：2025-06-25
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：9( 119-127 )