| Abstract | This dataset comprises combined gradient maps from nine pretrained convolutional networks for use with Multi-Targeted Gradient Training (MTGT). The combined gradients were optimized for a cyan (488 nm) Wavelength-Specific Map Attack. Each example corresponds to an original ImageNet-1K training or validation image, and the resulting combined gradient is stored as a single-channel (1 X 224 X 224) map. The combined gradient was computed by averaging the cross-entropy gradient of each network with respect to the true class, multiplying channel-wise by the RGB translation of a 488 nm cyan laser, and summing across channels. The ensemble includes SqueezeNet 1.0, SqueezeNet 1.1, GoogLeNet, AlexNet, VGG-11, VGG-11 with batch normalization, VGG-13, VGG-16, and VGG-16 with batch normalization, all sourced from the PyTorch Model Library and trained using a common preprocessing scheme in which the shortest side of the input was resized to 256 pixels while preserving its aspect ratio, followed by a 224 X 224 center crop. The networks were chosen to enable MTGT to better approximate the gradient behavior of architectures without residual connections. |
