Cycle GAN-Based Attack on Recaptured Images to Fool both Human and Machine

Recapture can be used to hide the traces left by some operations such as JPEG compression, copy-move, etc. However, various detectors have been proposed to detect recaptured images. To counter these techniques, in this paper, we propose a method that can translate recaptured images to fake “original images” to fool both human and machines. Our method is proposed based on Cycle-GAN which is a classic framework for image translation. To obtain better results, two improvements are proposed: (1) Considering that the difference between original and recaptured images focuses on the part of high frequency, high pass filter are used in the generator and discriminator to improve the performance. (2) In order to guarantee that the images content is not changed too much, a penalty term is added on the loss function which is the L1 norm of the difference between images before and after translation. Experimental results show that the proposed method can not only eliminate traces left by recapturing in visual effect but also change the statistical characteristics effectively.