A PCA-plus-local-differential-privacy framework for agricultural data sharing trades about 5.33% average accuracy loss for privacy on real datasets

It proposes a privacy-preserving mechanism combining PCA dimensionality reduction with Laplacian-noise local differential privacy to protect farmers' sensitive agricultural data and prevent reconstruction of original datasets. Existing agricultural data platforms (such as Farm2Facts, Local Line, and Open Food Network) collect vast agricultural data but lack privacy-preserving collaboration mechanisms; this work combines dimensionality reduction and differential privacy for agricultural collaborative research scenarios. The paper provides a formal security theorem and proof reducing the framework's security to that of the ε-LDP mechanism, showing computational indistinguishability between real-world and ideal-world views under a PPT semi-honest adversary model; it also evaluates robustness against membership inference attacks via power analysis on real datasets.

The framework identifies and groups farmers with similar characteristics through K-Means clustering and nearest-neighbor algorithms, supporting collaborator recommendations among farmers and between researchers and farmers. Prior studies focus heavily on securing smart devices from external attacks and rarely provide privacy-preserving collaboration frameworks tailored to small and medium-sized diversified farms; this work adds collaborator discovery and network formation on top of privacy protection. It presents K-Means clustering results on PCA-transformed data for the Wisconsin Farmer's Market dataset and the Crop Recommendation dataset, and simulates experiments by dividing the crop recommendation dataset into five distributed farmer's markets and a global dataset.

Researchers can train personalized models on identified collaborators' data via federated learning, or train models directly on aggregated privacy-protected data, with accuracy comparable to centralized training. The framework lets researchers train machine learning models without direct access to farmers' private data, since all analyses use only the differentially private outputs submitted by farmers. On the Wisconsin Farmer's Market dataset, Logistic Regression, Naive Bayes, and SVM achieve accuracies of 99.6%, 98.3%, and 99.5% on centralized data versus 92.1%, 93.8%, and 95.5% on aggregated privacy-protected data, an average accuracy loss of 5.33%; figures also show federated learning model accuracy versus the ε privacy budget.

It evaluates the privacy-utility trade-off via the power metric (fraction of correctly identified noisy samples) at a 5% false positive rate, determining optimal ε values for each model. It combines power analysis for privacy evaluation with utility evaluation across multiple classifiers, providing a basis for choosing ε values that balance privacy protection and data usability. On the Crop Recommendation dataset, the optimal ε values are 25, 35, and 35 for Logistic Regression, Naive Bayes, and SVM, respectively; corresponding power and accuracy analyses are also performed on the Farmer's Market dataset.

The framework targets farmers and researchers who need to share sensitive agricultural data for collaborative research, and is designed for deployment as a web-based sandbox environment on cloud platforms, including operations such as data aggregation, clustering, and filtering. It suits researchers identifying collaborators for a target farmer and training personalized models via federated learning, as well as farmers seeking collaborators based on farm-attribute similarity. The framework operates under assumptions that farmers are trusted, that researchers and the server are honest-but-curious, and that parties do not collude, and it applies to settings where data collection is relatively consistent and of high quality.

The paper notes that the computational requirements of differential privacy techniques may pose adoption challenges for resource-constrained farms, and that the framework relies on consistent and high-quality input data, so its effectiveness may be limited in regions where data collection is inconsistent or less accessible. Future work directions include improving scalability and adaptability, integrating decentralized data-sharing mechanisms such as blockchain, and incorporating real-time data processing. In addition, the loaded text is an incomplete version, and specific values in some figures (such as Figures 6 and 7 on federated learning accuracy versus ε) are not fully presented in the text, so readers needing precise privacy-utility curves should consult the original figures.