Analysis and Design of GPGPU-based Secure Visual Secret Sharing (VSS) Schemes

Abstract

Visual Secret Sharing (VSS) stands for sharing confidential image data across the par- ticipants as shares. As each share is not self-sufficient to compromise the secrecy, it is also called a shadow. Only the authorized participants can successfully recover the secret information with their shadows based on the type of VSS schemes. Recent advancements in VSS schemes have reduced the computation cost of encryption and decryption of the images. Still, real-life applications can not utilize these techniques with resource-constrained devices as they are computationally intensive. Another well- identified problem with VSS modalities is that they result in the distorted quality of the recovered image. The solution to improve the contrast of the decrypted image demand additional extensive processing. Therefore an ideal remedy is to exploit the advance- ments in hardware like GPGPU to propose novel parallel VSS schemes. The existing VSS schemes reconstruct the original secret image as a halftone image with only a 50% contrast. The Randomized Visual Secret Sharing (RVSS) scheme overcomes this disadvantage by achieving contrast of 70% to 90% for noise-like and 70% to 80% for meaningful shares (Mhala et al. 2017). But RVSS is computationally expensive. As a remedy we presented a GPGPU-based RVSS (GRVSS) technique that harness the high-performance computing power of the General-Purpose Computation on Graphics Processing Unit (GPGPU) for the data-parallel tasks in the RVSS. The presented GRVSS achieved a speedup range of 1.6× to 1.8× for four shares and 2.63× to 3× for eight shares, for different image sizes. To further enhance the visual quality of the retrieved image, we presented an effec- tive secret image sharing with super-resolution that leverage the deep learning super- resolution technique. The presented model outperformed the benchmark model in many-objective parameters with the recovered secret image contrast of 96.6% - 99.8% for noise-like and 64% - 80 % for meaningful shares. We used GPGPU to conserve the training time of the neural network model to achieve a speedup of 1.92× over the counterpart CPU super-resolution. Also, we presented an effective VSS model with super-resolution utilizing a Convo- lution Neural Network (CNN) intending to increase both the contrast of the recovered image and the speedup. The objective quality assessment proved that the presented model produces a high-quality reconstructed image with the contrast of 97.3% - 99.7% for noise-like and 78.4% - 89.7 % for meaningful shares having the Structural Simi- larity Index (SSIM) of 89% - 99.8% for the noise-like shares and 71.6% - 90% for the meaningful shares. The presented technique achieved an average speedup of 800× in comparison with the sequential model. The application of VSS schemes to medical images poses additional challenges due to their inherent poor quality. Also, medical imaging demands real-time VSS schemes to save the life of patients. We extended these concepts to propose a Medical Im- age Secret Sharing (MISS) model that deems fit for the Computer-Aided Diagnostic (CAD) tools. The presented fused feature extractor with a random forest classifier demonstrated that our VSS scheme reconstructs the secret medical images suitable for CAD tools. The result analysis confirmed the high-performance of the MISS with a 99.3% contrast and a 98% SSIM of the reconstructed image. MISS achieved an aver- age speedup of 800× in comparison with the sequential model. We used cluster sizes of 300, 500, and 1000 to obtain the MISS model’s CAD performance measures such as accuracy, sensitivity, specificity, precision, and F-measure using a presented fused feature extractor and a random forest classifier. The achieved precisions of 99.45%, 99.83%, and 100% for 300, 500, and 1000 cluster sizes prove the presented model’s suitability for the CAD systems.