ATTENTION-ENHANCED DEEP NEURAL PREDICTOR FOR HIGH-CAPACITY RDHEI

Abstract

The growing adoption of cloud computing, telemedicine, military communication, foren sic imaging, and big-scale multimedia storage has really pushed the need for secure and reverse information embedding mechanisms. Reversible Data Hiding in Encrypted Images (RDHEI) basically lets auxiliary data be embedded into encrypted images, but it also makes sure you can do exact recovery for both the hidden data AND the original image. That kind of exactness matters a lot in sensitive areas where any distortion in the image is basically not allowed, and confidentiality needs to stay protected during transmission, as well as storage. In this thesis report, we introduce an attention-enhanced deep neural prediction frame work aimed at high-capacity RDHEI in grayscale images. The whole idea is focused on a Convolu tional Block Attention Module (CBAM)-enhanced convolutional neural network predictor that is combined with an Adaptive Mean Predictor (AMP). The proposal is in spired by earlier asymmetric CNN-based and adaptive mean predictor approaches for RD HEI, in which a two-stage prediction strategy is used along with multi-MSB substitution to raise the embedding capacity. It is also guided by high-precision CNN prediction studies that indicate attention mechanisms can improve prediction accuracy. Compared to tradi tional local linear predictors and earlier CNN predictors that only provide a small amount of adaptive refinement when handling extracted features, the proposed CBAM-enhanced predictor tries to do better by applying both channel attention and spatial attention. The goal here is to highlight informative grayscale texture features and also bring out the im portant spatial structures more clearly. The framework you proposed really does a two-stage prediction, plus embedding, thing. First, you take a grayscale image and split it into Circle and Triangle sets, using a kind of checkerboard-style partitioning, which is a bit like alternating blocks. Then the Circle set becomes the contextual input to predict the Triangle set, with a CBAM-enhanced CNN predictor, or so it goes. Next, AMP steps in to predict the Circle set, via adaptive directional neighbourhood estimation, pretty much like it “looks around” in a tailored way. After that, the prediction error distributions from these stages are turned into embeddable locations, label maps, and auxiliary information. Those pieces then get used for reversible data hiding. The actual embedding process leans on multi-MSB substitution, label-map management, secure encryption, and separable data extraction plus image recovery. In the thesis, you’ll iv see the math setup for CBAM attention, how prediction errors are generated, how label maps are built, plus encryption, embedding, and the recovery procedures. So the proposed framework is supposed to be used for evaluation on the usual grayscale benchmark datasets like BOSSBase, BOWS-2, and UCID. For the comparison, it uses met rics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), prediction error variance, Embedding Rate (ER), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and also security metrics like entropy and correlation. Since we cannot just make up any additional unreported numerical experimental results, this report uses placeholders for the final CBAM-related result tables and graphs, even though we still show the baseline values that are already given in the reference literature. In short, the proposed grayscale RDHEI framework offers a high-capacity, kind of secure, and reversible solution. It is aimed at things like medical imaging, forensic communication, encrypted cloud storage, and privacy-preserving multimedia systems.