Here, we would like to train one PPRD for each output class
Below 10 figures are corresponding to training process of 10 one-PPRD model
Output class 0
Output class 1
Output class 2
Output class 3
Output class 4
Output class 5
Output class 6
Output class 7
Output class 8
Output class 9
Architecture for one-PPRD
Evaluation (use ground-truth PPRD)
Evaluation (use f(adv_x) PPRD)
