A Deep-Learning-Based Dual-Key Encryption Scheme With Comprehensive Performance Enhancement for Internet of Things Security

This paper presents a dual-key image encryption scheme based on deep neural network. First, a 4D memristive hyperchaotic map is constructed by integrating the Sine map, Logistic map, and a sinusoidal discrete memristor. The resulting map provides high Lyapunov exponents, a broad parameter domain, and an expanded chaotic range, making it suitable for cryptographic applications. On this basis, an image encryption scheme with comprehensive performance enhancements is designed by incorporating SqueezeNet and dual-key mechanism. The proposed scheme maintains a simplified structure while achieving strong performance in security, robustness, and efficiency, and its application in Internet of Things (IoT) environments is illustrated. Simulation and numerical results show that it achieves near-ideal performance in terms of NPCR, UACI, information entropy, and pixel correlation. Additionally, it exhibits strong resistance to high-intensity cropping attacks and multiple types of noise, and requires only 0.1856 s to encrypt a 512×512 color image. These results verify that the scheme ensures a high level of security and reliability in image transmission.