Strengthening the Transferability of Adversarial Examples Using Advanced Looking Ahead and Self-CutMix

Reconstructing images with attention generative adversarial network against adversarial attacks

“…In training dataset, this paper randomly selects 1000 normal images as target samples from 1000 classes in ImageNet dataset, 72 with one image selected from each class, and 15,000 adversarial examples are generated by 3 models: inception_v3, 73 inception_v4, 74 and resnet152, 14 using 5 adversarial attack algorithms: DIM, 75 FMFT-10, 76 LI-FGSM, 77 SIM, 78 and TIM 79 . Random transformations are applied to the input images at each iteration and diverse input patterns are created to improve the transferability of adversarial examples in DIM.…”

Boosting the Transferability of Ensemble Adversarial Attack via Stochastic Average Variance Descent

“…Nesterov iterative method [16] accelerated the craft speed of adversarial examples by applying accelerated gradients into the attack algorithm. The lookahead iterative method [17] tuned the update direction by recording the gradient in multiple previous steps to get rid of suboptimal regions during the update of the adversarial examples. These gradient-based attacks make the generated adversarial perturbations more accurate by optimizing the gradient, which is highly effective in both white-box and black-box attacks.…”

Improving the Robustness of Transformer-based Large Language Models with Dynamic Attention

“…Most adversarial examples are inherently unstable. Previous studies [23], [65], [28] have shown that adversarial examples experience a trade-off between transferability and imperceptibility, i.e., the imperceptible adversarial examples generated from the surrogate model can hardly fool the target model. To show that adversarial texts also have low transferability, we first train two models from the same training dataset, generate adversarial texts from one model, and then transfer them to another.…”