ROI-aware 4D hyperchaotic encryption for fundus image privacy with Lyapunov-verified security analysis
Abstract
Retinal fundus images transmitted through teleophthalmology networks carry protected health information subject to HIPAA and equivalent regulations, yet existing chaos-based encryption schemes apply uniform computational effort regardless of the clinical significance of image regions. This paper presents ROI-Aware 4D Chaos-Based Encryption (ROHE-4D), a privacy-preserving framework that allocates computationally stronger encryption operations to a selected anatomical ROI. A lightweight deterministic morphological segmentation procedure identifies the selected anatomical ROI, comprising the optic disc and macular region, without covering approximately 16.6% of image area on average, without model training. ROI pixels undergo key-dependent affine byte substitution with XOR diffusion with cipher feedback, whereas background pixels receive single-pass XOR diffusion, an architectural design choice. For each image, the keystream seed is derived from a 256-bit secret key, the SHA-256 image fingerprint, and a unique 128-bit nonce using a two-stage SHA-256 key-derivation construction. The uniformly generated secret key provides a 256-bit key space. The 4D system has one positive Lyapunov exponent (λ 1 ≈ 0.69) and the system is classified as chaotic. Morphological ROI segmentation achieves optic disc IoU of 0.752 ± 0.157 on synthetic images with circular ground-truth masks. Experiments on APTOS 2019 (3662 images, 512 × 512) and IDRiD (516 images, 1024 × 1024) demonstrate the following results: ciphertext entropy 7.9970 and 7.9989 bits/byte respectively. NPCR 99.62% and 99.61%, UACI 33.48% and 33.47%, adjacent-pixel ciphertext horizontal correlation 0.0186 (APTOS) and 0.0113 (IDRiD), key sensitivity ≈ 99.59% per 1-bit key perturbation, encryption latency 9.8 ms on CPU and 47.1 ms on Jetson AGX Orin. An analytical peak-memory estimate of approximately 2.5 MB based on principal array allocations at 512 × 512 and bit-exact lossless recovery across all 4178 test images.
