Adversarial Patch Detection for Artificial Intelligence Agriculture Disease Diagnosis Under Handheld Camera Capture: A Computer Certified-Detection Evaluation

Handheld-camera disease diagnosis is the most plausible route by which smallholder farmers could use artificial-intelligence classifiers, yet the same capture conditions that make it convenient also make it attackable: a printed patch held beside a leaf can flip a diagnosis without altering the lesion. the present analysis treats certified adversarial-patch detection as a design problem for that deployment setting and characterises the operating characteristics of a detection certificate under a declared variance structure. The analysis is a replication-based evaluation of a before-after-control-impact estimator, not an observational study: no field images, farms or respondents are used or claimed. Between-cluster standard deviation was fixed at 6.0, residual standard deviation at 4.5 and the intraclass correlation at 0.640, and power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. Fifty-four clusters were required to detect a 5-percentage-point effect at 80% power. The results indicate that the certified-detection framing converts patch defence from an unverifiable claim about a classifier into a measurable property of a capture protocol, and that the cluster counts realistic for a district-level deployment sit close to the detection boundary.