An Ultra-Lightweight Visual Privacy Protection System for Deep Optical Light Field Imaging Applications

Abstract

Light field imaging captures both spatial and angular information from a scene, thus enabling advanced computational photography applications such as post-capture refocusing, depth estimation and 3D reconstruction. Such capabilities make light field data extremely valuable for applications in areas such as healthcare, intelligent surveillance, virtual/augmented reality and robotics. However, 4D light field data is rich and very high dimensional, and it poses enormous threats to visual privacy, especially in sensitive personal or medical information scenarios. Classic encryption schemes are generally unable to handle the unique structure and bulk of the data streams that light fields generate in an efficient manner. In order to address this novel and unreleased threat, this thesis proposes PRESHMAC-256, a hybrid encryption scheme that combines the ultralightweight Present block cipher with the cryptographic robustness of HMAC-SHA256. The suggested method is able to encrypt the sub-aperture images extracted from light field captures with 64-bit blocks and with a 128-bit symmetric key derived securely from SHA-256 hashing. The encryption is designed to be lightweight and reversible, claiming a robust level of security with minimum computational and power overhead-an essential requirement for real-time and resource-constrained applications like embedded or mobile devices. The experimental validation carried out using the EPFL Light Field dataset to prove that the presented approach indeed works. Evaluation metrics considered are histogram analysis, information entropy, pixel correlation coefficients, PSNR, SSIM, key sensitivity analysis, occlusion attack resilience and avalanche effect measurements. The findings emerge favoring PRESHMAC-256 as not only ensuring good levels of image security and fidelity, but also decreasing model complexity and processing latency as compared to the traditional encryption schemes. Certainly, this work thus opens possibilities for such an application on the grounds of a fairly practical, highly scalable and secure environment in which data privacy and computation efficiency would both matter significantly.