Convolutional neural network-based high capacity predictor estimation for reversible data embedding in cloud network

Document Type : Research Paper


Research Scholar, Department of Computer Science, Chirashree Institute of Research and Development (CIRD), University of Mysore, Karnataka, India.


This paper proposes a reversible data embedding algorithm in encrypted images of cloud storage where the embedding was performed by detecting a predictor that provides a maximum embedding rate. Initially, the scheme generates trail data which are embedded using the prediction error expansion in the encrypted training images to obtain the embedding rate of a predictor. The process is repeated for different predictors from which the predictor that offers the maximum embedding rate is estimated. Using the estimated predictor as the label the Convolutional neural network (CNN) model is trained with the encrypted images. The trained CNN model is used to estimate the best predictor that provides the maximum embedding rate. The estimation of the best predictor from the test image does not use the trail data embedding process. The evaluation of proposed reversible data hiding uses the datasets namely BossBase and BOWS-2 with the metrics such as embedding rate, SSIM, and PSNR. The proposed predictor classification was evaluated with the metrics such as classification accuracy, recall, and precision. The predictor classification provides an accuracy, recall, and precision of 92.63\%, 91.73\%, and 90.13\% respectively. The reversible data hiding using the proposed predictor selection approach provides an embedding rate of 1.955 bpp with a PSNR and SSIM of 55.58dB and 0.9913 respectively.


Main Subjects