Research Article

Protecting Privacy in Shared Photos via Adversarial Examples Based Stealth

Figure 9

An intuitive understanding of adversarial images for detection task in the high dimensional space. (a) Different cases that feature point moves between the be-object class and not-object class in the high dimensional feature space. (b) Different cases that feature point moves among different specific classes. Each subspace with a color represents a specific class. The subspace in the be-object region but not belonging to any specific class represents its score of belonging to any class which is lower than our set threshold.
(a)
(b)