基于 JPEG 图像的隐写方法通常利用原始图像的边信息来提高隐写安全性． 然而 现有的 JPEG 图像隐写方法在使用边信息时需要提前获取未压缩的原始图像，当原始图 像不可获得时，往往不能把边信息用于设计隐写方法． 文中提出了一种在不具有未压缩的 原始图像的情况下，通过估计图像的边信息来提高 JPEG 图像隐写安全性的方法． 首先， 使用卷积神经网络对 JPEG 载体图像进行恢复，得到估计的原始图像; 其次，通过得到的 估计图像计算取整误差;然后利用取整误差来控制嵌入过程中离散余弦变换( DCT) 系数 的修改方向;最后结合 DCT 系数的失真代价和 STC 编码来实现隐写嵌入． 文中提出的边 信息估计方法能够应用于现有的 JPEG 图像自适应隐写方法，如 JC-UED 和 J-UNIWAＲD． 实验结果表明，该方法能较好地恢复图像的边信息，并能显著提高现有 JPEG 图像自适应 隐写方法的安全性． 

JPEG steganography usually takes advantage of the side information of raw images to improve steganographic security． However， the raw images are indispensable for existing JPEG steganography to apply side information． When the raw images are not available， the side information cannot be used to design more secure steganography． In this paper，we propose to improve the steganographic security by estimating the side information from JPEG images when the raw images are not available． Firstly， the convolution neural networks are used to restore the estimated spatial images from JPEG images． Secondly，the rounding errors are calculated with the estimated images． Then， the rounding errors are utilized to control the modification directions of DCT coefficients in the embedding process． Finally，by integrating the selected modification directions and the distortion costs of DCT coefficients， the STC-based steganographic embedding is implemented． The proposed method can apply to existing adaptive JPEG steganographic schemes，such as JC-UED and J-UNIWAＲD． The experimental results show that the proposed method can effectively estimate the side information from JPEG images and can improve the security of existing adaptive JPEG steganographic schemes．