A Proofs

The case when N 3 can be trivially generalized using mathematical induction. We assume the bias terms are zeros without loss of generality, i.e., In this section, we discuss the unstable Pearson's correlation for small variance inputs, i.e., It generates perturbed samples that maximize both cross entropy and regularizer. Jacobian matrix of the logits vector and p is the probits of the model. In this section, additional visualizations are provided in Figure 1 and Figure 1 to demonstrate that IGR improves attribution robustness. The original and adversarial images from different datasets are shown in the first two columns.