F-MIFGSM: adversarial attack algorithm for the feature region

The current mainstream attack algorithms IFGSM and MI-FGSM will carry out global perturbation to the original example when they carry out adversarial attack. F-MIFGSM algorithm is proposed to solve the problem of poor attack concealment. This algorithm extracts the output information of the convolutional layer of the neural network, marks out the feature region of the original example through the deconvolution network and mapping algorithm, and uses the MI-FGSM algorithm to carry out the specific region attack. In the experiment, 60,000 training sets in MNIST data set are used to train the self-constructed convolutional neural network, and 10,000 test sets in the data set are attacked, and the generated adversarial examples are input into the neural network for classification prediction. The experimental results show that the F-MIFGSM algorithm can significantly reduce the disturbance range to the original example and reduce the defects between the adversarial example and the original example under the condition of increasing the attack success rate.