6740526b78c0b230e41ae61d8ca07cf5-Supplemental.pdf

This supplementary material includes two sectionsi.e., details of baselines and descriptions of experiments. Mutualinformation measures the essential correlation between twosamples and can effectively estimate the similarity between featureszand input datax. The resultswith differentγ values are shown in Figure 1, where Figure 1(a) are original samples, Figure 1(b) are generated samples by clean features, and Figures 1(c)-1(f) are generated samples from the perturbed features with differentγ.