An Embedded Cost Learning Framework Based on Cumulative Gradient Rewards

Weixuan Tang, Yingjie Xie

Abstract: The structure of the Generative Adversarial Network (GAN) has demonstrated good performance in various tasks, mainly comprising two competing sub-networks. The GAN has the potential to effectively generate artificial samples that closely resemble the actual sample distribution. The field of steganography utilizing the Generative Adversarial Network (GAN) structure has witnessed a wealth of research with highly successful outcomes. This paper proposes a steganography framework that integrates reinforcement learning and introduces a new reward function to analyze the embedding cost of images in the steganography problem. In this framework, the reward function assigns distortion values to each pixel of the image and relates the security performance of steganography. Based on the conducted experiments, an enhanced steganographic embedding scheme can ultimately be achieved.