With the rapid advancement of agricultural digitalization, crop disease detection has become critical for ensuring food security and improving agricultural productivity (Elijah et al., 2018; Cornia, 1985). In many remote or underdeveloped regions, due to the lack of professional expertise and detection equipment, farmers often transmit crop images to external agricultural centers for manual or automated analysis (Baldi and La Porta, 2020). However, existing methods frequently struggle to balance model accuracy with data privacy protection. In addition, the detection systems must operate efficiently on resource-constrained devices typical of rural environments, while ensuring secure handling of sensitive data during transmission and storage. Consequently, the development of efficient, automated, and privacy-preserving disease detection systems that are both lightweight and reliable is crucial for promoting smart agriculture.

In recent years, deep learning-based object detection algorithms have achieved remarkable progress in plant disease recognition (Senthil Pandi et al., 2024; Ravì et al., 2017; Zhang et al., 2017). Lightweight models, particularly the YOLO series, have attracted considerable attention for their fast detection speeds and high accuracy (Zhang et al., 2024; He et al., 2024; Liu et al., 2022). Sangaiah et al. (2024) proposed T-YOLO-Rice, based on YOLOv4, to improve small-target detection such as rice leaf spots, outperforming YOLOv7 but remaining limited to a single task. To address diverse diseases and class imbalance, Roy and Bhaduri (2022) developed Dense-YOLOv4 by integrating DenseNet and an enhanced PANet, achieving 96.20% mAP and 93.61% F1-score for mango disease detection, and demonstrating generalization to grape and tomato diseases. Lin et al. (2023) built YOLO-Tobacco based on YOLOX-Tiny by adding HMU and CBAM modules, improving outdoor tobacco leaf detection (80.56% AP, 69 FPS), although its adaptability to multiple diseases remains limited. Building upon these advances, Li et al. (2023a) introduced MG-YOLO, integrating multi-head self-attention, BiFPN, and GhostCSP modules, achieving 98.3% accuracy at 0.009 seconds per image and surpassing YOLOv5 by 6.8% in complex environments. In addition to single-task detection, recent studies have explored joint detection and tracking paradigms (Li et al., 2023b, 2024), leveraging reinforcement learning to achieve object recognition and continuous tracking in dynamic environments. For example, Li et al. (2025) proposed a reinforcement learning-based joint detection and tracking paradigm for compact HFSWR target detection and tracking, which effectively improves detection probability and tracking performance.

In image tasks related to object detection, image enhancement has also emerged as an important research direction in recent years. For example, researchers have proposed a reinforcement learning-based human visual perception-driven image enhancement method (Luo, 2024). Liu et al. (2025a) introduced a framework that cascades an aerial image enhancement module with AC3Net, while Xiao et al. (2024) proposed a neuromorphic computing-based underwater image enhancement network (UIEN), which simulates visual system perception and employs unsupervised learning to address multiple types of underwater image degradation and validate its effectiveness. Despite these significant advances in image enhancement, most existing studies still overlook data privacy issues, as unencrypted images transmitted over networks are vulnerable to theft or misuse. This further highlights the necessity of integrating image encryption with recognition.

Therefore, with increasing emphasis on data privacy, researchers have begun integrating image encryption with disease detection to achieve end-to-end security without compromising performance. Qin et al. (2014) proposed SecSIFT, a method that performs SIFT feature extraction directly within the encrypted domain, effectively safeguarding sensitive image data while maintaining high detection accuracy and computational efficiency. Building on this idea, Man et al. (2021) integrated convolutional neural networks with chaotic encryption, enabling intelligent privacy protection for both image and text data, and laying the foundation for secure image processing in agriculture. Kumar et al. (2021) introduced SP2F, a privacy-preserving framework combining blockchain and deep learning, with a two-level privacy engine and stacked LSTM networks to improve UAV data authentication and resilience. Furthermore, Kethineni and Gera (2023) proposed an IoT security model that integrates sparse capsule autoencoders and attention-based GRUs for lightweight detection and data protection, achieving 99.9% accuracy and F1 score, highlighting its potential for agricultural data security.

Our work aims to develop a lightweight deep learning model for precise crop disease detection and robust image-level privacy protection. Optimized for resource-constrained edge devices, it ensures real-time, high-precision identification of various disease types. Additionally, to secure data transmission, we integrate a novel image encryption scheme based on a 3D chaotic cube, effectively preventing unauthorized access without compromising detection performance. Our model has been comprehensively evaluated on real-world datasets and outperforms existing methods in detection accuracy, computational overhead, and privacy protection. This solution offers a practical and secure pathway for smart agriculture applications. Our approach addresses two key challenges in plant disease detection: data privacy and detection accuracy.

Our main contributions are as follows:

The rest of this paper is organized as follows: Section 2 reviews related work. Section 3 details the encryption method. Section 4 introduces the detection model. Section 5 describes data processing and optimization. Section 6 presents experiments and analysis. Section 7 concludes the paper.Section 8 highlights the system’s significance, practical value, and limitations.

Traditional encryption algorithms such as AES and DES are inadequate for real-time protection of agricultural images due to high dimensionality, redundancy, and computational overhead associated with such data. Although deep learning has achieved success in plant disease detection tasks, most existing studies overlook privacy concerns during image transmission and processing. To contextualize the proposed integrated system, this section reviews key image encryption techniques and plant disease detection approaches.

In eggplant disease detection, image encryption is essential for data security by preventing unauthorized access, tampering, and maintaining integrity. Ciphertext transmission enhances system security and reduces the risk of cyberattacks. This section presents a novel image encryption method based on a 3D Chaotic Cube Encryption Scheme, which consists of four steps: preparation of the key and image, generation of 3D key and index array, permutation encryption and XOR operation, and save the encrypted image and key. Compared to frequency- and chaos-based methods (Jui-Cheng and Guo, 2000; Armand Eyebe Fouda et al., 2014; Jammula et al., 2022), the proposed scheme offers stronger resistance to attacks and superior performance. Figure 2 illustrates the encryption framework, and Figure 3 shows the original and encrypted eggplant images.

Deep learning-based object detection has demonstrated remarkable performance in disease recognition, with the YOLO series widely used for its speed and accuracy (Dai et al., 2022; Wang et al., 2024; Xie et al., 2024). However, traditional YOLO models struggle when dealing with small objects and class imbalance (Obu et al., 2023; A.N. and A.P., 2022). To address these issues, we propose an enhanced lightweight model, YOLOv11n-12D. As shown in Figure 5 , the architecture consists of four components: Stem, Backbone, Neck, and Head. In our model, YOLOv11n serves as the student model, while YOLOv12s acts as the teacher. By leveraging knowledge distillation and detection loss, we enhance recall, reduce missed detections, and maintain efficiency, making it suitable for large-scale agricultural applications. The distillation process is detailed in Figure 6 .

The following are the steps of the distillation process:

Since logits vary significantly in magnitude, a temperature parameter T = 4.0 is applied to smooth them, stabilize gradients, and generate soft labels for distillation.

When T = 1.0, the distribution reduces to standard softmax, limiting the ability to learn from the teacher. We adopt Kullback-Leibler divergence as the distillation loss: Distillation Loss = T ² × KL(p _t ∥ p _s ). Here, T ² offsets the gradient scaling effect caused by smoothing. A smaller KL value indicates better alignment between the student and teacher outputs, KL as defined in Equation 27:

After each epoch,the validation set is evaluated using mAP, Precision, and Recall to monitor the effectiveness of the distillation strategy on student model performance.

To ensure high efficiency and accuracy in eggplant disease detection, a large-scale dataset was constructed, encompassing four categories: eggplant rot, fruit borer, healthy samples, and thrips. All images were annotated in YOLO format with standardized bounding boxes and precise class labels. The annotations were reviewed by agricultural experts to ensure high-quality and consistent labeling.The dataset was collected from multiple eggplant cultivation bases, encompassing various growth cycles, diverse lighting conditions, and all developmental stages of pest/disease infestation (from initial infection to characteristic symptom manifestation). Specifically, the test set comprises 745 representative images (containing 1516 annotated instances), while the remaining 7520 images were partitioned into training (5264 images) and validation (2256 images) sets at a 7:2:1 ratio. This scientifically designed partitioning scheme ensures both sufficient training data volume and reliable evaluation of model generalization capability. Sample differences between healthy and diseased eggplants are shown, highlighting the visual variability between categories ( Figure 7 ). To address the limited quantity and variable quality of the collected raw images, we employed the Albumentations data augmentation library to enhance dataset diversity and improve model generalization and robustness (Buslaev et al., 2020; Korra et al., 2022). The overall workflow for material preparation and optimization is illustrated ( Figure 8 ), and the specific data augmentation strategies are detailed as follows:

The applied augmentation techniques significantly enhance the model’s robustness and generalization, allowing more reliable recognition of eggplant disease features under diverse conditions. Furthermore, the integration of Focal Loss and CloU Loss improves detection accuracy, achieving a final accuracy of 99.1% and effectively reducing the miss detection rate, thereby improving applicability in real-world agricultural scenarios.

This section presents the evaluation metrics used to assess the performance of the two core components of the proposed system: the 3D chaotic cube-based encryption scheme for image security, and the YOLOv11n-12D-based detection model for eggplant disease diagnosis.

This paper proposes an integrated system for eggplant disease detection that combines image encryption and deep learning-based recognition. The system employs a lightweight encryption scheme based on 3D chaotic mapping and pixel permutation to secure image transmission with low computational overhead. It then utilizes an optimized YOLOv11n-12D model to process the decrypted images, achieving high detection accuracy and real-time performance. A teacher–student knowledge distillation strategy is incorporated to further enhance model robustness. Experimental results demonstrate that the system not only safeguards data privacy but also outperforms existing methods in accuracy, speed, and stability, offering a reliable solution for smart agriculture. At the same time,our future research will focus on enabling disease detection directly on encrypted images to eliminate the risk of data leakage during decryption. This will involve exploring privacy-preserving techniques like homomorphic encryption and designing lightweight models that can operate effectively in the encrypted domain.

Our study proposes a framework that integrates image encryption with deep learning based object detection for real time, privacy-preserving crop disease monitoring. The designed 3D chaotic cube encryption scheme demonstrates strong security performance, achieving high entropy (7.6195), low pixel correlation (0.0084), and strong resistance to statistical and differential attacks (NPCR = 99.63%, UACI = 32.85%). Meanwhile, the YOLOv11n-12D model retains the detection performance of the teacher model while achieving fast inference speed (2.7 ms), with a notable mAP improvement of +6.5% in small-object detection such as eggplant thrips. This solution offers a promising approach for advancing smart agriculture in rural or resource limited areas. By encrypting images before transmission and decrypting them only during model inference, the framework strikes a practical balance between data security and operational efficiency. Its compatibility with edge devices further supports deployment in real world scenarios, where data privacy, bandwidth limitations, and low computing resources are common challenges. Despite the promising results, the current framework still requires decryption before detection, which introduces a temporary risk of data exposure. Future work will focus on privacy preserving deep learning techniques that support inference directly in the encrypted domain, such as homomorphic encryption or secure multi party computation. Further validation on larger and more diverse crop datasets is also needed to assess generalization. Enhancing the interpretability of both the detection model and the encryption process will help improve transparency and user trust in practical applications.