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．