Benchmarking Evaluation of Byzantine-Robust Federated Aggregation for Computer Agriculture Disease Classifiers Under Adversarial Client Participation: An Artificial Intelligence Contamination Bound for Diagnostic Integrity

Federated learning for crop-disease classification across heterogeneous farm clients is vulnerable to Byzantine participants who submit corrupted model updates, yet the aggregation literature reports robustness without a design-level account of how many adversarial clients a given federation can tolerate before diagnostic integrity collapses. The analysis treats the aggregation rule as a statistical estimator of the honest update direction and characterises the breakdown point of coordinate-wise median and trimmed-mean aggregation under a declared contamination model. A before-after-control-impact design is used to benchmark power across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with 2000 replications per design cell and an intraclass correlation of 0.640. The contamination bound predicts that coordinate-wise median aggregation tolerates up to 50% adversarial clients in the limit but loses diagnostic integrity at a threshold that depends on the between-cluster standard deviation of 6.0 and the residual standard deviation of 4.5. The benchmarking results indicate that 54 clusters are required to detect a 5 percentage point degradation at 80% power. The contribution is a formal contamination bound that links aggregation-rule choice to the minimum federation size needed to certify diagnostic integrity, together with a benchmarking protocol that makes that link auditable.