Image steganography aims to conceal the secret information inside another carrier image. And by embedding the information into the carrier image, the carrier image may suffer certain image distortion. Thus, not only the hiding algorithm should be carefully designed, but also the carrier image should be meticulously selected during the hiding process. This paper follows the idea of creating suitable cover images instead of selecting the ones by presenting a unified architecture which combines real-time object detection based on convolutional neural network, local style transfer using generative adversarial network and steganography together to realize real-time carrier image generation. The object in the carrier image is first detected using a fast object detector and then the detected area is reconstructed through a local generative network. The secret message is embedded into the intermediate generated images during the training process in order to generate an image which is suitable as an image carrier. The experimental results show that the reconstructed stego images are nearly indistinguishable to both human eyes and steganalysis tools. Furthermore, the whole carrier image generation process with GPU implementation can achieve around 5 times faster than the regular CPU implementation which meets the requirement of real-time image processing.

中文翻译：

---

基于生成对抗网络和快速目标检测的实时图像载体生成

图像隐写术旨在将秘密信息隐藏在另一个载体图像中。并且通过将信息嵌入到载体图像中，载体图像可能遭受某些图像失真。因此，不仅应该精心设计隐藏算法，而且在隐藏过程中也要精心选择载体图像。本文提出了创建合适的封面图像而不是选择封面图像的想法，提出了一种统一的体系结构，该体系结构结合了基于卷积神经网络的实时对象检测，使用生成对抗网络的局部样式传递和隐写术，共同实现了实时载体图像代。首先使用快速物体检测器检测载体图像中的物体，然后通过本地生成网络重建检测到的区域。秘密消息在训练过程中被嵌入到中间生成的图像中，以便生成适合作为图像载体的图像。实验结果表明，重建后的隐身图像几乎无法与人眼和隐写分析工具区分开。此外，采用GPU实施的整个载波图像生成过程可以比常规CPU实施快约5倍，从而满足了实时图像处理的需求。