Attention-Guided Digital Adversarial Patches on Visual Detection

Figure 5

In this model, images are transformed into tensor form as the input of the visualization network. The final heatmap is obtained by feature extraction of each layer and weight. By modifying the last layer of network structure, the requirement of input image size is reduced.