Steganography in Style Transfer

Steganography entails concealing secret data within a given medium for covert communication. In recent years, style-transferred images have been widely disseminated on social media, offering a novel multimedia carrier for steganography. However, there is currently a lack of steganographic techniques specifically designed for style-transferred images. In this paper, we propose disguising the steganographic tool as a deep neural network (DNN) performing style transfer tasks. In our method, a neural network is manipulated to transfer the style of a given image to a target style, while also embedding secret data into the given image. Meanwhile, a trained receiving network is used to extract the embedded data. The same pre-trained network used by the processing network and the receiving network matches the feature maps of secret data at the same layers. Under the guidance of secret data, a stego image is generated after being trained by the processing network and the receiving network an appropriate number of times, and secret data can be extracted from stego image. Experimental results demonstrate that our method is more effective and secure compared to existing steganographic algorithms that achieve data embedding by modifying image content.