Fully Connected Neural Network Based Carrier Estimation Mechanism on Encrypted Images for Data Hiding in Cloud Network

The work proposes a fully connected neural network (FCNN) based approach for detecting the Carrier blocks for embedding the data in encrypted images in the cloud network. In a data embedding process, the determination of non- carrier pixels that provide underflow and overflow during the data embedding process plays a major role. The location map for the non-carrier blocks is usually compressed and embedded in the encrypted image along with the hidden data. The embedding rate and peak signal-to-noise ratio (PSNR) are limited due to the storage of huge location map information on the image. Therefore, the proposed approach uses the FCNN network to detect the Carrier blocks /non-carrier blocks which highly minimizes the additional location map information to be embedded. In the embedding phase, a trained FCNN network is utilized to detect the carrier blocks, in which the FCNN network is trained with the labels that are generated by trial 0’s and 1’s embedding process. Two approaches are utilized in training the FCNN that includes FCNN with predictor only (FCNN-PO) and FCNN with sub-block fully (FCNN-SF) schemes in detecting the carrier blocks. In the extraction phase, the same FCNN model is used to detect the carrier blocks from which the data and actual encrypted image can be reconstructed. The performance of the carrier detection process was evaluated using measures such as precision, recall, and accuracy, while the data hiding process was evaluated using measures such as structural similarity index measurement (SSIM), PSNR, and embedding rate. The FCNN-PO carrier/non-carrier classification process results in an average accuracy of 98.59% in detecting the carrier while providing an SSIM, PSNR, and embedding rate of 0.9926, 58.86db and 1.97bpp respectively during the data embedding process when evaluated using the Bows-2 dataset.