关键词: 攻击, 防御, 图像放缩, 图像切割, 像素值统一化

Abstract: In order to defend against existing attacks on artificial intelligence algorithms (especially artificial neural networks) as much as possible, while reducing the additional costs, the Rattan algorithm based on examples preprocessing was proposed. By cutting the unimportant information part of the image, unifying the neighboring pixel values and image scaling, the examples were preprocessed to destroy the adversarial disturbance and generate new examples that pose less threat to the model, so as to ensure high accuracy of model recognition. Experimental results show that the Rattan algorithm can defend against some adversarial attacks against MNIST, CIFAR10 datasets and neural network models such as squeezenet1_1, mnasnet1_3 and mobilenet_v3_large with less overhead than similar algorithms, and the minimum accuracy of the defended examples can reach 88.50%; meanwhile, it will not reduce the examples accuracy too much when processing clean examples, and the defense effect and defense cost are better than the comparison algorithm.

Key words: attacks, defense, image scaling, image cutting, pixel value normalization