Experiment 8: Potential Method 4 & Integrated LPs judgement (adv_attack=i_FGSM, y/y'=y', model=CNN)
Note that LP_i = B if B_log_prob_i > log_prob_diff_line_i
LP_1, LP_2, and LP_3 are LPs for the convolutional layers; LP_4 is the LP for the first ReLU layer.
As shown in the following figures, it is difficult to tell that Potential Method 4 bring any improvement based to Potential Method 3.
LP_i risk score distribution with delta=0.1 (i={1, 2, 3, 4})
LP_i risk score distribution with delta=0.2 (i={1, 2, 3, 4})
LP_i risk score distribution with delta=0.3 (i={1, 2, 3, 4})
LP_i risk score distribution with delta=0.4 (i={1, 2, 3, 4})
LP_i risk score distribution with delta=0.45 (i={1, 2, 3, 4})
