Agricultural visual recognition: a novel privacy-preserving framework based on 4D chaos and Mamba networks

With the advancement of smart agriculture and Internet of Things (IoT) based monitoring systems, secure and accurate crop disease recognition for large-scale agricultural management remains challenging. Existing methods mainly focus on recognition performance while insufficiently considering data privacy protection and computational efficiency, which limits their deployment in resource-constrained agricultural environments. To address these issues, this study proposes a privacy-preserving crop disease recognition framework that combines multi-image encryption and lightweight detection. First, a four-dimensional chaotic encryption scheme is developed to secure agricultural image transmission by incorporating chaotic sequence generation, multi-channel pixel permutation, and DNA-based diffusion operations. The proposed encryption method improves data security and anti-attack capability while maintaining low computational cost. Furthermore, a lightweight Mamba-LSDet network is designed for crop disease recognition after secure image transmission and decryption. Equipped with a cross-stage gated module for adaptive feature extraction and a shared prediction head for parameter reduction, the network achieves accurate and efficient disease detection. Experimental evaluations demonstrate the effectiveness of the proposed encryption scheme in terms of security and efficiency. Moreover, detection experiments after image decryption show that Mamba-LSDet achieves mAP@50 and mAP@50-95 values of 0.841 and 0.505, respectively, outperforming the lightweight YOLOv11n model. The proposed framework provides an effective solution for secure and lightweight crop disease recognition in agricultural IoT applications.