Integrating saliency maps with CNN and chaos theory for robust image scrambling and descrambling

Abstract Images are becoming increasingly prevalent in various aspects of our daily lives. It’s understandable that sensitive fields like healthcare, defense, and surveillance rely on capturing images to extract more detailed information. This research stems from the growing need for an adaptive image encryption solution that remains secure during transmission. Recent studies indicate that systems incorporating adaptive elements and cognitive intelligence can lead to innovative advancements in image security. Our hybrid model, which combines chaos-based cryptography with deep learning, presents a promising solution for protecting image data. By leveraging and improving pattern learning, aiding in the reconstruction process. The goal of this proposed model is to establish a framework for image scrambling and descrambling, where the unpredictability of chaos intertwines with the adaptability of deep networks. The result for PSNR and SSIM values between original and scrambled images is ranging from (7.34 to 18.50 dB), and SSIM values (0.007 to 0.176) respectively, which indicates the strong distortion and high security. Regarding the retrieval process, a perfect reconstruction is achieved across the entire dataset. With metrics reaching unity for SSIM and zero for MSE, the data suggests a complete absence of information loss. The robustness analysis confirms that the scrambled image maintains high integrity under different attacks and correlation values. The proposed work also includes an ablation study to evaluate the contribution of individual components and to quantify the performance improvements achieved through their combined integration.