基于注意力机制的高容量通用图像隐写模型

随着深度学习与隐写技术的发展,深度神经网络在图像隐写领域的应用越发广泛,尤其是图像嵌入图像这一新兴的研究方向.主流的基于深度神经网络的图像嵌入图像隐写方法需要将载体图像和秘密图像一起输入隐写模型生成含密图像,而最近的研究表明,隐写模型仅需要秘密图像作为输入,然后将模型输出的含密扰动添加到载体图像上,即可完成秘密图像的嵌入过程.这种不依赖载体图像的嵌入方式极大地扩展了隐写的应用场景,实现了隐写的通用性.但这种嵌入方式目前仅验证了秘密图像嵌入和恢复的可行性,而对隐写更重要的评价标准,即隐蔽性,未进行考虑和验证.提出一种基于注意力机制的高容量通用图像隐写模型USGAN,利用注意力模块,USGAN的编码器可以在通道维度上对秘密图像中像素位置的扰动强度分布进行调整,从而减小含密扰动对载体图像的影响.此外,利用基于CNN的隐写分析模型作为USGAN的目标模型,通过与目标模型进行对抗训练促使编码器学习生成含密对抗扰动,从而使含密图像同时成为攻击隐写分析模型的对抗样本.实验结果表明,所提模型不仅可以实现不依赖载体图像的通用嵌入方式,还进一步提高了隐写的隐蔽性.

High-capacity Universal Image Steganographic Model Based on Attention Mechanism

With the development of deep learning and steganography, deep neural networks are widely used in image steganography, especially in a new research direction, namely embedding an image message in an image. The mainstream steganography of embedding an image message in an image based on deep neural networks requires cover images and secret images to be input into a steganographic model to generate stego-images. But recent studies have demonstrated that the steganographic model only needs secret images as input, and then the output secret perturbation is added to cover images, so as to embed secret images. This novel embedding method that does not rely on cover images greatly expands the application scenarios of steganography and realizes the universality of steganography. However, this method currently only verifies the feasibility of embedding and recovering secret images, and the more important evaluation criterion for steganography, namely concealment, has not been considered and verified. This study proposes a high-capacity universal steganography generative adversarial network (USGAN) model based on an attention mechanism. By using the attention module, the USGAN encoder can adjust the perturbation intensity distribution of the pixel position on the channel dimension in the secret image, thereby reducing the influence of the secret perturbation on the cover images. In addition, in this study, the CNN-based steganalyzer is used as the target model of USGAN, and the encoder learns to generate a secret adversarial perturbation through adversarial training with the target model so that the stego-image can become an adversarial example for attacking the steganalyzer at the same time. The experimental results show that the proposed model can not only realize a universal embedding method that does not rely on cover images but also further improves the concealment of steganography.