Fig. 3

The block diagram of the entire process of generating ally patches and readjustment of the classification output. The ally patch extractor intrinsically extracts 'promising’ ally patches from the input image. Then identical, yet independent, copies of a pre-trained convolutional deep network are used to evaluate the classification of each of the ally patches. Finally, the fusion stage decides the most appropriate final label of the entire image. Any possible adversarial patch is counter-attacked during the ally patch extraction stage. Fled adversarial patches from the evaluator are taken care of in the final stage