Adaptive chaotic image encryption model driven by deep neural network

Abstract

The growing demand for secure multimedia transmission requires image encryption schemes that not only provide strong statistical security but also exhibit adaptive robustness against evolving attacks. This paper proposes ACE-DNN, an adaptive chaotic image encryption model driven by a deep neural network, in which chaotic control parameters are dynamically predicted from plaintext statistical descriptors. Unlike conventional chaotic encryption schemes, ACE-DNN embeds the neural predictor into the chaotic control loop, introducing a data-dependent nonlinear mapping that enhances trajectory divergence and phase space complexity. The hybrid logistic–sine chaotic system achieves an average Lyapunov exponent of 0.9715–0.9827, ensuring rapid exponential separation of trajectories and stronger diffusion. Experimental evaluation on 512× 512 grayscale images demonstrates high statistical security, with ciphertext entropy averaging 7.9973 bits, residual redundancy below 0.003 bits per pixel, and near-zero adjacent pixel correlations. Differential attack metrics indicate number of pixels change rate of 99.634% and unified average changing intensity of 33.487%, closely matching theoretical expectations. The effective key space, accounting for floating-point precision and neural network weight hashing, exceeds 2256, providing strong resistance against brute-force attacks while increasing the complexity of known-plaintext cryptanalysis through adaptive parameter generation and image-dependent chaotic evolution. Furthermore, the framework maintains computational feasibility, with encryption completed in 0.183 s per image. This work highlights the theoretical and practical advantage of integrating deep learning with adaptive chaotic dynamics, offering a scalable and secure paradigm for modern multimedia encryption.