Adversarial Patch Detection for Artificial Intelligence Agriculture Disease Classifiers Under Cooperative Field Deployment: A Computer Certified-Radius Evaluation

Adversarial patches pose a distinctive threat to convolutional disease classifiers deployed across cooperatively managed smallholder fields, because a single patch placed in a shared scouting route can corrupt predictions for many plots at once. The framework defines a certified detection radius as the smallest perturbation magnitude at which the smoothed detector provably flags an adversarial input, and it derives a lower bound on that radius in terms of the classifier's Lipschitz constant, the smoothing variance and the abstention threshold. A replication-based power analysis of the before-after-control-impact design used to validate the detector was conducted across 2000 design cells with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.64; 52 clusters were required to detect a 5 percentage-point effect at 80% power. The analysis indicates that certified detection is feasible for cooperative deployment only when cluster counts reach 24 or more, and that the intraclass correlation of 0.64 dominates the design's sensitivity to patch-induced error. The contribution is a formal certificate that converts a classifier's Lipschitz geometry into an auditable detection guarantee suitable for cooperative field governance.