Evaluation of Concept-Activation Pathway Audits for Artificial Intelligence Agriculture Disease Classifiers Under Adversarial Image Perturbation: An Interpretability Robustness Bound for Smallholder Field Deployment Safety

Concept-activation pathway audits are increasingly proposed as a safety check for vision-based crop-disease classifiers before smallholder field deployment, yet the audit's own robustness under adversarial image perturbation has not been characterised. the present analysis treats the audit as a statistical estimator of a classifier's concept reliance and derives a finite-sample bound on how far the audited concept-activation profile can drift when the input image is perturbed within an ε-ball. The bound is expressed as a Lipschitz-type inequality linking the audit's sensitivity to the classifier's local Lipschitz constant, the perturbation radius, and the number of concept probes. 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 required 54 clusters to detect a 5-percentage-point effect at 80% power. The results indicate that the audit's robustness bound widens sharply as the classifier's Lipschitz constant grows, and that the before-after-control-impact design recovers the audit's drift signal only when cluster counts exceed the 54-cluster threshold. The paper concludes that concept-activation pathway audits can support smallholder deployment safety only when paired with an explicit robustness bound and a cluster count sufficient to detect the drift the bound predicts.