Deep neural networks are extremely vulnerable to externalities from intentionally generated adversarial examples which are achieved by overlaying tiny noise on the clean images.However,most existing transfer-based att...Deep neural networks are extremely vulnerable to externalities from intentionally generated adversarial examples which are achieved by overlaying tiny noise on the clean images.However,most existing transfer-based attack methods are chosen to add perturbations on each pixel of the original image with the same weight,resulting in redundant noise in the adversarial examples,which makes them easier to be detected.Given this deliberation,a novel attentionguided sparse adversarial attack strategy with gradient dropout that can be readily incorporated with existing gradient-based methods is introduced to minimize the intensity and the scale of perturbations and ensure the effectiveness of adversarial examples at the same time.Specifically,in the gradient dropout phase,some relatively unimportant gradient information is randomly discarded to limit the intensity of the perturbation.In the attentionguided phase,the influence of each pixel on the model output is evaluated by using a soft mask-refined attention mechanism,and the perturbation of those pixels with smaller influence is limited to restrict the scale of the perturbation.After conducting thorough experiments on the NeurIPS 2017 adversarial dataset and the ILSVRC 2012 validation dataset,the proposed strategy holds the potential to significantly diminish the superfluous noise present in adversarial examples,all while keeping their attack efficacy intact.For instance,in attacks on adversarially trained models,upon the integration of the strategy,the average level of noise injected into images experiences a decline of 8.32%.However,the average attack success rate decreases by only 0.34%.Furthermore,the competence is possessed to substantially elevate the attack success rate by merely introducing a slight degree of perturbation.展开更多
基金Fundamental Research Funds for the Central Universities,China(No.2232021A-10)Shanghai Sailing Program,China(No.22YF1401300)+1 种基金Natural Science Foundation of Shanghai,China(No.20ZR1400400)Shanghai Pujiang Program,China(No.22PJ1423400)。
文摘Deep neural networks are extremely vulnerable to externalities from intentionally generated adversarial examples which are achieved by overlaying tiny noise on the clean images.However,most existing transfer-based attack methods are chosen to add perturbations on each pixel of the original image with the same weight,resulting in redundant noise in the adversarial examples,which makes them easier to be detected.Given this deliberation,a novel attentionguided sparse adversarial attack strategy with gradient dropout that can be readily incorporated with existing gradient-based methods is introduced to minimize the intensity and the scale of perturbations and ensure the effectiveness of adversarial examples at the same time.Specifically,in the gradient dropout phase,some relatively unimportant gradient information is randomly discarded to limit the intensity of the perturbation.In the attentionguided phase,the influence of each pixel on the model output is evaluated by using a soft mask-refined attention mechanism,and the perturbation of those pixels with smaller influence is limited to restrict the scale of the perturbation.After conducting thorough experiments on the NeurIPS 2017 adversarial dataset and the ILSVRC 2012 validation dataset,the proposed strategy holds the potential to significantly diminish the superfluous noise present in adversarial examples,all while keeping their attack efficacy intact.For instance,in attacks on adversarially trained models,upon the integration of the strategy,the average level of noise injected into images experiences a decline of 8.32%.However,the average attack success rate decreases by only 0.34%.Furthermore,the competence is possessed to substantially elevate the attack success rate by merely introducing a slight degree of perturbation.
