Question about the paper: Does Instance Selective Whitening have an information leakage problem？

[Slonna]

Because photometric noise augmentation is used in the generation process of the mask matrix, and whether this operation may implicitly model data in the target interference domain and then get the performance gain?