AN ML-ENHANCED FRAMEWORK FOR PIXEL-LEVEL SELECTIVE IMAGE ENCRYPTION

Abstract

In an era dominated by digital communication, ensuring the secure and efficient transmission of images is critical for applications such as telemedicine, forensics, and surveillance. Traditional full-image encryption approaches, while secure, impose significant computational overhead, making them impractical for real-time or resource- constrained environments. This paper introduces a Machine Learning (ML)-Enhanced Selective Image Encryption (SIE) framework that performs pixel-level adaptive encryption using a pre-trained sensitivity estimation model. The proposed framework dynamically applies AES-256-GCM encryption to only the most sensitive regions, guided by a continuous sensitivity map, while leveraging PBKDF2 for secure key derivation. A conditional full- encryption strategy is invoked to eliminate residual information leakage in high-risk cases. Experimental results on BSDS500, CelebA, and NIH Chest X-ray datasets demonstrate that the proposed method achieves superior entropy (>7.69), NPCR (>58.98%), and UACI (>19.98%) values, with a 70% reduction in encryption time compared to full AES-GCM. The results confirm that the proposed framework provides a robust, cryptographically standardized, and computationally efficient solution for real-time secure image transmission.