Information Hiding of Specific Image Region Using Adversarial Examples

Deep neural networks have been widely used in our daily life. However, malicious organizations may collect image data and identify users for commercial purposes. Even worse, criminals may steal private information through image recognition to achieve illegal purposes. As well known, deep neural networks are easily attacked by adversarial examples. We propose a range oriented adversarial examples generation algorithm to fool AlexNet by improving the non-directional DeepFool algorithm. Furthmore we combine it with object detection and propose an information hiding algorithm of specific image region. Experimental results show that compared with traditional methods such as mosaic, blur, and partial occlusion, our scheme is more concealed while ensuring the usability of the image.