DCT-CNN Hybrid Model for High-Capacity and Secure Data Concealment in Encrypted Images

This research article presents a revolutionary approach to address the challenges of reversible data embedding in encrypted images within cloud networks. Our pioneering algorithm combines the potency of the Discrete Cosine Transform (DCT) and Convolutional Neural Network (CNN) to introduce a novel DCT-CNN hybrid model, specifically designed for efficient and high-capacity reversible data hiding. Diverging from conventional methods, our approach eliminates the need for trial data embedding during predictor selection. The initial phase involves estimating the optimal predictor using CNN, where training labels are forged through a trial embedding procedure. The selected predictor, providing the highest embedding rate, is assigned as the class label for the corresponding encrypted image. Users secure their images through a three-tier encryption process before transferring them to the cloud. During the testing phase, the trained CNN model processes encrypted images uploaded by users to determine the predictor class with the maximum embedding capacity. Data embedding is then executed using the classified predictor. The proposed predictor storage demonstrates impressive results, achieving an accuracy, recall, and precision of $93.9 \%, 92.6 \%$, and $95.3 \%$ in average, respectively, particularly for the BossBase dataset. Additionally, our reversible data embedding algorithm, employing the DCT-CNN paradigm, achieves a commendable embedding rate of 69.1bpp, along with an average PSNR of 38.8 dB and SSIM of 0.93 in average. Evaluations conducted on datasets, including BossBase and BOWS-2, convincingly demonstrate the stoutness and effectiveness of the proposed method. This research introduces a unique perspective to the realm of reversible data embedding in cloud networks, offering a promising solution for secure and high-capacity data concealment within encrypted images.