Publication

Abstract : With the rapid adoption of AI-driven decision systems in agriculture, manufacturing and IoT-based surveillance, protecting the privacy of sensitive visual data has become essential. Conventional chaos-based encryption methods treat all pixels uniformly, leading to unnecessary computation and insufficient protection of semantically important regions. This paper proposes a semantic-guided image encryption framework that integrates self-supervised Vision Transformers with chaotic systems to improve security for real-world intelligent systems. A DINOtrained ViT automatically generates attention maps to detect critical regions without labelled data. These maps guide global scrambling. It is followed by multi-level local scrambling, bitplane permutation, and multi-round feedback diffusion. Unlike existing schemes, the proposed work introduces attention-aware scrambling, ensuring stronger protection of meaningful regions while keeping the system computationally suitable for IoT, drone imaging, and industrial monitoring applications. Multiple chaotic systems-including logistic, sine, tent, Lorenz, and hybrid combinations-were evaluated, with the Lorenz system giving the best security performance NPCR =99.647%, UACI =33.586, entropy =7.998. The proposed ramework is suitable for securing image data in cyber-physical systems, including applications such as autonomous vehicles, smart surveillance, and IoT networks