Evaluation of Byzantine-Robust Federated-Learning Aggregation for Computer Agriculture Cooperative Pest-Monitoring Under Adversarial Sensor Poisoning: An Artificial Intelligence Convergence-Guarantee Audit

Cooperative pest-monitoring networks that aggregate sensor evidence across farms through federated learning face a structural threat: a minority of compromised nodes can submit adversarial updates that survive averaging and corrupt the global model. Byzantine-robust aggregation rules are widely proposed as the remedy, yet their convergence guarantees rest on assumptions about update geometry that agricultural sensor deployments rarely satisfy. the present analysis audits those guarantees. It formalises the aggregation problem for a cooperative pest-monitoring federation, states the standard Byzantine-robust estimators (coordinate-wise median, trimmed mean, Krum), and derives the conditions under which each retains a bounded convergence rate under adversarial sensor poisoning. The audit shows that the guarantees are conditional on a contamination fraction and a heterogeneity bound that are not observable in deployment, and that the effective contamination fraction is inflated by non-IID pest phenology across farms. A before-after-control-impact design is characterised under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) to show how many clusters a cooperative trial would need to detect a 5-percentage-point effect at 80% power: 52. Convergence-guarantee audits, not point accuracy comparisons, should govern the adoption of Byzantine-robust aggregation in computer agriculture.