Disrupting Image-Translation-Based DeepFake Algorithms with Adversarial Attacks

DeepNude, a deep generative software based on image-to-image translation algorithm, excelling in undressing photos of humans and producing realistic nude images. Although the software was later purged from the Internet, image translation algorithms such as CycleGAN, pix2pix, or pix2pixHD can easily be applied by anyone to recreate a new version of DeepNude. This work addresses the issue by introducing a novel aspect of image translating algorithms, namely the possibility of adversarially attacking these algorithms. We modify the input images by the adversarial loss, and thereby the edited images would not be counterfeited easily by these algorithms. The proposed technique can provide a guideline to future research on defending personal images from malicious use of image translation algorithms.