Smart Watermark to Defend against Deepfake Image Manipulation

Deepfake image manipulation has become a serious security threat to the social network. Currently, there are limited studies on protective methods that are against Deepfake image manipulation. To tackle this problem, we here propose an adversary attack based smart watermark model, which adds unperceptive watermarks to images so that the images become adversary examples to Deepfake models. When the Deepfake manipulates these watermarked images, the manipulated images become blur. The manipulation thus can be easily recognized by human and machines. Our experiments have shown that our model outperforms the SOTA and can be used to effectively prevent Deepfake manipulation.