Fig. 9
From: Attacking convolutional neural network using differential evolution
The distribution of perturbed pixels of successful attacks. The graphs show the locations of perturbed pixels of successful attack on 500 CIFAR-10 images (32×32 resolution) and 150 ImageNet images (227×227 resolution). The experiment is conducted on for four types of networks: all convolutional network (AllConv), network in network (NiN), VGG, and AlexNet BVLC, each under two DE settings: 0.5/0.1/0.1 and 0.1/0.1/0.1. The vertical and horizontal axes of the graph respectively indicates the length and width of the image