A Gradient analysis

To better understand why our generated confounder noise can make the data unlearnable, we can also gain some insights according to optimization gradient. Empirically, if one image provides a large gradient in a backpropagation, this image has a lot of learnable knowledge, and vice versa. Figure 9 shows the accuracy curves of our method during the training epoch. Then we give a detailed discussion about this setting. To better understand this adaptive setting, we first illustrate the assumption on the data owner's The model trainer wishes to train a denoiser against the noise generated by the ConfounderGAN.