Evaluation of Concept-Probe Leakage Audits for Computer Vision Agriculture Disease Classifiers Under Participatory Retraining: An Interpretability Confidentiality Bound for Farmer Data Sovereignty

Concept-probe audits are increasingly proposed as a way to certify that computer vision disease classifiers for agriculture do not encode sensitive farmer information. The problem is framed as a confidentiality bound: the probability that a probe recovers a protected attribute from the model's internal representation, conditional on the retraining regime and the audit's own design parameters. A formal treatment is developed in which the leakage signal is decomposed into a between-cluster component and a residual component, and the audit's operating characteristics are characterised under a declared variance structure with an intraclass correlation of 0.64, a between-cluster standard deviation of 6.0 and a residual standard deviation of 4.5. 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. The analysis indicates that 54 clusters are required to detect a 5 percentage point leakage effect at 80 per cent power, a requirement that exceeds the cluster counts typically available in participatory retraining programmes. Concept-probe leakage audits should be reported with an explicit confidentiality bound rather than a binary pass or fail, and that the bound should be treated as a condition of the data-sovereignty agreement under which farmer data are contributed.