Privacy-Preserving Medical Image Transmission Using Residual Autoencoder Representations and Lightweight Feature-Domain Encryption
Main article
Abstract
The expansion of telemedicine, mobile health platforms, and connected imaging devices has created a pressing need for image-protection techniques that defend patient confidentiality while remaining viable for processors with strict energy and memory budgets. Conventional block ciphers such as AES provide robust guarantees but were not designed for the redundancy patterns of medical imagery, and purely chaos-based proposals frequently shift complexity from encryption into key-scheduling. This study introduces a privacy-preserving transmission pipeline that performs encryption in the feature domain learned by a residual autoencoder, rather than over raw pixels. A compact latent tensor and a learned residual map are produced by the encoder, then independently protected by a two-pass column-oriented SHA3-256 stream cipher with cylindrical feedback. Evaluation on a composite medical dataset spanning chest radiographs, brain MRI, abdominal CT, and retinal photographs reports an average NPCR of 99.60%, UACI of 33.46%, and information entropy above 7.99 in the encrypted feature domain. Adjacent-pixel correlations are driven from above 0.95 to below 0.005, and the chi-square statistic drops by more than two orders of magnitude. The full encrypt-transmit-decrypt-reconstruct cycle completes in roughly 285 ms per 256×256 image on a single mid-range GPU, with the cryptographic stage alone taking 28 ms. Limitations of the design under noisy wireless channels are quantified and addressed through optional channel coding, which restores reconstruction SSIM above 0.85 at σ = 20. The framework offers a practical balance between confidentiality, fidelity, and computational lightness for medical image transmission in resource-constrained healthcare settings.
