基于深度强化学习的自适应图像隐写算法  

Adaptive Image Steganography Algorithm Based on Deep Reinforcement Learning

作　　者：钱清 龙永 蒋忠远 段春红 王宏 QIAN Qing;LONG Yong;JIANG Zhongyuan;DUAN Chunhong;WANG Hong(Key Laboratory of Blockchain and Fintech of Guizhou Provincial Colleges and Universities,School of Information,Guizhou University of Finance and Economics,Guiyang 550025,Guizhou,China;College of Big Data Statistics,Guizhou University of Finance and Economics,Guiyang 550025,Guizhou,China)

机构地区：[1]贵州财经大学信息学院贵州省高等学校区块链与金融科技重点实验室,贵州贵阳550025 [2]贵州财经大学大数据统计学院,贵州贵阳550025

出　　处：《计算机工程》2024年第8期319-327,共9页Computer Engineering

基　　金：国家自然科学基金(61902085);贵州省科技计划项目(黔科合基础-ZK[2021]一般312);贵州省教育厅青年科技人才成长项目(黔教合KY字[2021]132);贵州省教育厅高校创新团队项目(黔教技[2023]065号);贵州财经大学校级项目(2022ZXSY160)。

摘　　要：针对自适应图像隐写中如何在轻量化隐写、最佳嵌入定位、高隐匿输出三者之间实现均衡的问题,提出一种基于深度强化学习的自适应图像隐写算法(AISA-DRL)。设计一种轻量化安全隐写网络,在降低模型隐写成本的前提下加强模型对图像隐写特征的提取能力,增强载密图像的安全性和稳定性。首先将具有高效特征融合特性的EPSANet引入EfficientnetV2-s,得到改进的EPSA-EfficientnetV2-s,以提高像素级嵌入过程的表征能力,从而获得最优像素修改位张量。随后将秘密信息与最优像素修改位张量加权求和得到载密图像。最后通过学习隐写分析网络对载密图像进行最优像素级奖励分配,根据设计的最小化失真函数通过梯度回传来更新网络参数,以获得最佳嵌入位置,从而实现秘密信息的最佳化嵌入。实验结果表明,AISA-DRL算法的模型参数量减少了94.22%,FLOPs减少了24.88%,与其他基于强化学习的隐写方案相比,在不同经典隐写分析器下的检错率提高了2.48%~6.55%。此外,在不同载荷下生成的载密图像PSNR值均在30 dB以上,不仅提高了模型对像素修改位的定位准确率,而且使隐写网络具有更强的表征能力。This study proposes an Adaptive Image Steganography Algorithm based on Deep Reinforcement Learning(AISA-DRL)to realize a balance among lightweight steganography,optimal embedding positioning,and high hidden outputs.A lightweight secure steganography network is designed to enhance the ability of the model to extract stego features of an image based on the premise of reducing the steganography cost and optimizing the security and stability of the stego image.First,an EPSANet with efficient feature fusion characteristics is introduced into EfficientnetV2-s to obtain an improved EPSA-EfficientnetV2-s.This integration enhances the representation ability of the pixel-level embedding process to obtain the optimal tensor for the pixel modification bit.The stego image is then calculated using the weighted sum of the secret information and the optimal tensor of the pixel modification.Finally,by learning the steganographic analysis network,the optimal pixel-level reward is assigned to the stego image,and the network parameters are updated by gradient backpropagation according to the designed minimum distortion function.This process ensures that the best embedding position is obtained,and the optimal embedding of the secret information is realized.The experimental results show that the parameters of the AISA-DRL are reduced by 94.22%,and FLOPs are reduced by 24.88%.Compared with steganographic methods based on reinforcement learning,the error detection rate of the AISA-DRL increases by 2.48%-6.55%.Additionally,the PSNR values of the stego images generated under different loads are all greater than 30 dB.Thus,the proposed AISA-DRL improves the locational accuracy of the model for the pixel modification bit.The steganographic network of the AISA-DRL demonstrates a stronger characterization capability compared to other algorithms.

关 键 词：图像隐写 强化学习 隐写分析 嵌入策略 特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]