CNN-transformer fusion with multi-chaotic dynamics: a content-adaptive framework for high-security image and video encryption

The increasing adoption of connected imaging systems and edge devices has created a growing need for efficient and robust protection of image and video data against unauthorized access and cryptanalytic attacks. This paper presents a content-adaptive encryption framework that integrates convolutional neural networks (CNNs), Transformer-based feature extraction, multi-chaotic dynamics, and learned cryptographic state generation. The proposed architecture combines local spatial representations from CNNs with global contextual information captured by a Transformer encoder to construct a content-dependent latent representation. This representation is subsequently transformed into a high-dimensional cryptographic state that controls the initialization and selection of subsequent chaotic encryption operations. Unlike conventional approaches that treat the dimensionality of a floating-point latent representation as a direct measure of cryptographic key length, the proposed framework distinguishes learned cryptographic state generation from the effective discrete key space. The encryption process further combines multiple chaotic maps with permutation and diffusion mechanisms to improve statistical properties and resistance to common cryptanalytic attacks. Experimental evaluation on standard image and video data demonstrates near-uniform encrypted histograms, low adjacent-pixel correlations, high information entropy, and strong sensitivity to encryption parameters. The framework is also implemented on a Raspberry Pi to assess its computational feasibility on resource-constrained edge platforms. For a $$512\times512$$ RGB image, the complete processing time is 5.7984 s on the CPU and 14.1082 s on the Raspberry Pi, including cryptographic-state generation, encryption, and decryption. These results demonstrate the feasibility of deploying the proposed architecture on embedded edge platforms while identifying computational latency as an important consideration for future optimization. The proposed approach provides a content-dependent and hybrid deep-learning/chaos framework for securing image and video data in edge-oriented applications.