Detection Mechanisms of One-Pixel Attack

Algorithm 1 Algorithm of candidate detection method.

Input: an adversarial image generated by a one-pixel attack, a DNN classifier, and a label set with

Output: a set of pixels containing candidate victim pixels

1: Initialize model with the target DNN model

2: Randomly chose pixels from the image, and for each point, randomly set a color as present pixel set

3: Calculate the confidence of the image on all labels

4: for all each label do

5: while true do

6:  Calculate the change of the confidence on when modifying with parent pixels

7:  Generate offspring pixels based on Equation (5)

8:  Calculate the change of confidence on when modifying with offspring pixels

9:  Select pixels with the highest confidence changed as new parent pixels

10:  if all top confidences changed on targets are larger than then

11:   Save the top pixels as candidate victim pixels

12:   Set

13:   Break while loop

14:  end if

15:  if while loop is over times then

16:   Break while loop

17:  end if

18: end while

19: end for

20: if AttackSucc is true then

21: return candidate victim pixels

22: else

23: return fail to find candidates

24: end if

Algorithm 1 Algorithm of candidate detection method.