Evaluation of Concept-Embedding Memorization Audits for Computer Vision Agriculture Disease Classifiers Under Farmer Feedback Retraining: An Interpretability Privacy Bound for Proprietary Field Imagery

Farmer feedback retraining is the standard mechanism by which computer vision disease classifiers for smallholder agriculture are kept current, yet the same mechanism that improves accuracy also embeds proprietary field imagery into model parameters in ways that concept-probe audits cannot detect. The analysis is a replication-based evaluation of the estimator's operating characteristics, not an empirical study of field data: no observational data are used or claimed. Power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design requires 54 clusters to detect a 5 percentage point effect at 80% power. The paper derives a memorization-audit bound that relates concept-embedding leakage to the retraining feedback channel, shows that the bound is identifiable under the before-after-control-impact structure, and specifies the minimum detectable effect curve that governs audit design. The contribution is a methodological framework that makes the privacy cost of farmer feedback retraining auditable without requiring access to the proprietary imagery itself.