Abstract: Existing image steganography methods have made significant progress in concealment and antiattack capabilities, but still suffer from low image quality, limited information embedding capacity, and insufficient model stability and generalization ability. To address these issues, this paper proposes an image steganography model based on an improved GAN and selfdistillation (RCSDGAN). First, a residualchannel attention mechanism is designed and integrated into the DenseNet architecture to optimize its structure. Second, an encoding network and decoding network based on the enhanced DenseNet are constructed, combined with a discriminative network to form a complete image steganography framework. Finally, a selfdistillation training strategy is introduced. The selfdistillation loss is defined by calculating the difference between the teacher model’s output and the student model’s output from the same network, and this loss is incorporated into the overall loss function to enhance model stability and generalization capability. Experimental results demonstrate that, at an embedding rate of D=1bpp, the steganographic images achieve a PSNR of 50.4131dB, an SSIM of 0.9992, and an Accuracy of 99.96%.

Key words: image steganography; generative adversarial network; channel attention mechanism; residual connectivity; selfdistillation

摘要： 现有的图像隐写方法在隐蔽性与抗攻击性方面取得了一定进展，但仍存在生成图像质量较低、信息嵌入容量受限以及模型稳定性与泛化能力不足的问题.为此，提出一种基于改进GAN与自蒸馏的图像隐写模型(RCSDGAN).首先，设计了一种残差通道注意力机制，并将其集成到DenseNet架构中，实现了DenseNet结构的优化；其次，构建了一种基于改进DenseNet的编码网络与解码网络，并结合判别网络，形成了完整的图像隐写框架；最后，引入自蒸馏训练策略，通过计算教师模型与同一网络中学生模型输出之间的差异定义自蒸馏损失，并将其引入到总损失函数中，提升模型的稳定性与泛化能力.实验结果表明，当D=1bpp时，隐写图像的峰值信噪比(PSNR)达到50.4131dB，结构相似性(SSIM)达到0.9992，解码准确率(Accuracy)达到99.96%.

关键词: 图像隐写；生成对抗网络；通道注意力机制；残差连接；自蒸馏