Federated learning of crop-disease classifiers across quarantine-silenced phytosanitary archives: a computer-science evaluation of privacy-utility trade-offs for cross-border outbreak analytics

Abstract

Phytosanitary archives hold the labelled crop-disease observations that cross-border outbreak analytics require, yet quarantine regulation, commercial confidentiality and national data sovereignty keep those archives closed to centralised training. the present analysis formalises federated learning over such archives as a constrained optimisation problem in which each participating authority trains a local classifier on its own held-out data and shares only parameter updates under a differential-privacy mechanism. The framework is evaluated analytically rather than empirically: we derive a Rényi-differential-privacy accountant for the interleaved Gaussian mechanism, bound the generalisation gap of the federated averaged classifier as a function of archive heterogeneity and privacy budget, and characterise the privacy-utility frontier as a Pareto surface whose curvature is governed by the number of participating archives and the Dirichlet concentration of their label distributions. The analysis is illustrated with a reanalysis of NASA POWER MERRA-2 monthly air temperature for Kampala, Uganda (0.35° N, 32.58° E), 2000–2023, comprising 312 monthly observations with a mean of 22.443 °C and a linear warming trend of −0.1564 °C per decade, used here only to fix the climatic covariate distribution under which the classifier operates. The contribution is a decision procedure that lets a regional plant-protection organisation choose a privacy budget, an archive-participation set and a communication round count that together meet a stated outbreak-detection utility target without centralising any raw observation.