A CycleGAN Adversarial Attack Method Based on Output Diversification Initialization

The powerful image generation capabilities of generative adversarial networks (GAN) bring great threats to applications related images. Style transfer realize the style transform between domains through which we can easily modify images like portraits and calligraphy. To eliminate negative impact caused by forged images, there emerged technical methods detect might trigger remedial actions afterwards but cannot prevent maliciously tampered content from spreading over network media. Therefore, some scholars put forward idea protecting hostile with attack. However, initial random noise perturbation be effectively mapped output space. In order improve visual effect attacks, this paper proposes an attack algorithm based on diversification initialization (ODI) for CycleGAN. We firstly utilize find effective starting point attack, then use Project Gradient Descent (PGD) iteratively modifying loss function. Experimental results demonstrate that introduction ODI enlarge distance original output. It achieves better in identifying generated targeted model, does not significantly increase disturbance, guarantee normal