Benchmarking Evaluation of Federated Distillation for Computer Agriculture Disease Classifiers Under Heterogeneous Edge-Device Label Noise: An Artificial Intelligence Privacy-Utility Bound for Cooperative Surveillance

Federated distillation promises to train plant-disease classifiers across edge devices without centralising raw leaf imagery, yet its privacy-utility behaviour under heterogeneous label noise remains uncharacterised for cooperative agricultural surveillance. The analysis treats the before-after-control-impact estimator as the evaluation instrument and characterises its operating characteristics under a declared variance structure with between-cluster standard deviation 6.0, residual standard deviation 4.5 and intraclass correlation 0.64. Power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The bound separates a distillation-consistency term, a noise-heterogeneity term and a privacy-budget term, and predicts that utility loss grows with the dispersion of device-level noise rather than with its mean. The design analysis indicates that 56 clusters are required to detect a 5 percentage point effect at 80 percent power under the declared variance structure. The contribution is a benchmarkable privacy-utility bound that converts label-noise heterogeneity into a design constraint for cooperative surveillance, together with an evaluation protocol that makes the bound testable without observational data.