Neural Networks Implemented by Differential Evolutionary Algorithms to Counter Attacks

It is urgent and necessary to investigate the adversarial attacks on different models, the attack patterns and attack methods of the adversarial attacks. In this paper, three convolutional neural network models, LeNet, ResNet and DenseNet, were used to train image recognition for the Cifar-10 multispecies dataset, and a differential evolutionary algorithm was used to implement a counterattack on the neural network. Among them, the Drop-out mechanism and Batch-Nomalization layer were added to the neural network model to solve the overfitting problem and improve the gradient dispersion problem of the neural network, respectively, and finally the differential evolution algorithm was used to achieve the attack on the neural network model. The experimental results have shown that the image recognition accuracies of LeNet, ResNet, and DenseNet models reached 53.83%, 92.95%, and 93.17%, respectively. When the differential evolutionary algorithm was used to implement the adversarial attack on the three neural network models, 93%, 78%, and 69% were achieved, respectively. Comparing the attack success rate of the three network models, it can be found that the result is consistent with the image recognition rate and network structure robustness of the three models.