基于全局-局部扰动协同的对抗样本增强算法OA

To enhance the diversity,local adaptability,and transferability of perturbations in in-put transformation attacks,a dynamic global-local adaptive perturbations DGLAP)framework is pro-posed.This framework integrates three core modules to generate and optimize adversarial perturba-tions.The global block shuffling GBS)module reorganizes input information using cross-scale and random diffusion strategies to extract model-invariant features.The local adaptive perturbation LAP module applies diverse transformations to sensitive image regions based on dynamic region partitio-ning and edge-continuity constraints.Additionally,the dynamic weighted random walk DWRW mechanism adaptively adjusts transformation weights through a stochastic strategy that balances ex-ploration and exploitation.Experimental results on the ImageNet dataset demonstrate that DGLAP out-performs baseline methods in terms of attack success rate on mainstream models,such as ResNet18 and ResNet101,and exhibits superior transferability against adversarially trained models.