Backdoor Trigger Neutralization for Artificial Intelligence Agriculture Disease Classifiers Under Cooperative Model Sharing: A Computer Poisoning-Recovery Evaluation

Cooperative model sharing lets agricultural cooperatives pool disease-classifier training without pooling raw leaf imagery, but it also lets a single poisoned contributor install a backdoor trigger that survives aggregation and fires on-field. the present analysis formalises trigger neutralisation as a recovery problem and evaluates a before-after-control-impact (BACI) estimator for measuring neutralisation efficacy under a declared variance structure. No observational data are used or claimed: the analysis characterises the operating characteristics of the design across 2000 replications per cell, with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.640. Cluster counts of 8, 16, 24 and 32 were crossed with effect sizes of 3, 5 and 8 percentage points. The design attains 80% power at 54 clusters for a 5-percentage-point effect. Estimated power rose monotonically with effect size and cluster count, and the minimum detectable effect at 80% power fell as clusters increased. The contribution is a design-level account of what a poisoning-recovery evaluation can and cannot detect, together with the boundary conditions under which a null neutralisation result is informative rather than merely underpowered.