Scholarlink

<p>Tomato (Solanum lycopersicum) leaf disease detection in smart agriculture faces critical challenges in balancing accuracy with data privacy, as traditional centralized deep learning approaches and conventional federated learning methods remain vulnerable to gradient leakage, model inversion, and poisoning attacks that can expose commercially sensitive agricultural data including proprietary crop traits and regional farming patterns. This study proposes and evaluates a novel homomorphic encryption-enhanced federated learning (HEFL) framework specifically designed for privacy-preserving agricultural image processing, implementing dual-layer encryption at both client and server sides to enable secure model training without exposing sensitive information during any stage of the learning process. We conducted comprehensive experiments using nine different deep learning architectures (ResNet50, EfficientNet, DenseNet121, MobileNetV2, and others) on tomato leaf disease classification tasks to assess both performance and security robustness of the proposed framework. Our results demonstrate that MobileNetV2 achieved the highest test accuracy within the HEFL framework while maintaining computational efficiency suitable for resource-constrained agricultural environments, and the framework successfully preserved model performance comparable to centralized approaches while providing significant protection against inference and reconstruction attacks. We conclude that HEFL represents a viable and scalable solution for privacy-preserving agricultural disease detection that addresses the critical need for secure, decentralized crop monitoring systems. The primary limitations of this study include the computational overhead introduced by homomorphic encryption operations and the need for further validation across diverse agricultural datasets and real-world deployment scenarios, which can be addressed through optimized encryption schemes and extensive field testing in future work.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Privacy-aware plant disease detection: federated learning with homomorphic encryption on image data

  • 期刊论文

  • 发表日期：

    2025年12月6日

查看全文
  • Md Zahin Muntaqim,
  • Hasan Muhammad Kafi,
  • Tangin Amir Smrity

摘要

Tomato (Solanum lycopersicum) leaf disease detection in smart agriculture faces critical challenges in balancing accuracy with data privacy, as traditional centralized deep learning approaches and conventional federated learning methods remain vulnerable to gradient leakage, model inversion, and poisoning attacks that can expose commercially sensitive agricultural data including proprietary crop traits and regional farming patterns. This study proposes and evaluates a novel homomorphic encryption-enhanced federated learning (HEFL) framework specifically designed for privacy-preserving agricultural image processing, implementing dual-layer encryption at both client and server sides to enable secure model training without exposing sensitive information during any stage of the learning process. We conducted comprehensive experiments using nine different deep learning architectures (ResNet50, EfficientNet, DenseNet121, MobileNetV2, and others) on tomato leaf disease classification tasks to assess both performance and security robustness of the proposed framework. Our results demonstrate that MobileNetV2 achieved the highest test accuracy within the HEFL framework while maintaining computational efficiency suitable for resource-constrained agricultural environments, and the framework successfully preserved model performance comparable to centralized approaches while providing significant protection against inference and reconstruction attacks. We conclude that HEFL represents a viable and scalable solution for privacy-preserving agricultural disease detection that addresses the critical need for secure, decentralized crop monitoring systems. The primary limitations of this study include the computational overhead introduced by homomorphic encryption operations and the need for further validation across diverse agricultural datasets and real-world deployment scenarios, which can be addressed through optimized encryption schemes and extensive field testing in future work.