Benchmarking Evaluation of Adversarial Patch Defenses for Computer Agriculture UAV Crop-Scouting Under Field-Deployable Threat Models: An Artificial Intelligence Certified-Radius Bound for Spray-Recommendation Integrity

This is the authors' abstract. We don't add key points for this paper.

Adversarial patches pose a credible threat to unmanned aerial vehicle (UAV) crop-scouting systems that convert imagery into spray recommendations, yet the field lacks a certified guarantee that a defended classifier's recommendation remains within a bounded distance of the clean-input recommendation. The framework formalises the spray-recommendation pipeline, defines a certified radius over the patch-perturbation ball, and derives a bound that relates the certified radius to the Lipschitz constant of the recommendation head and the margin of the defended classifier. A replication-based evaluation of the before-after-control-impact estimator was conducted under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells at cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design achieved 80% power at 52 clusters for a 5-percentage-point effect. The certified-radius bound provides a deployable integrity certificate for spray recommendations under field-realistic patch threat models, and the benchmarking protocol exposes the trade-off between certified radius and recommendation coverage that field deployments must resolve.