Adversarial example generation method based on normal distribution function optimization iterative gradient

Adversarial attack refers to the destruction of deep learning models by adding perturbations to the image. At present, there remains room for improvement in black box attacks, particularly when dealing with the models equipped with defense mechanisms. Our propose a new attack way for making confrontation sample using the optimal iterative gradient of normal distribution functions. In each iteration, the sampled image is predicted several times from the vicinity of the current image, and their gradient information is calculated. Then, the correction gradient is calculated based on the normal distribution function, and this correction gradient is used for correction in the next iteration. This iterative approach ensures stability in the update direction, reduces overfitting, improves the generalization ability of adversarial examples. We conducted various analyses on the ImageNet1K dataset^[7] to prove the usefulness of our method. The results showed that our method has improved the probability of successful attacks