Privacy-Preserving Vertical Federated Learning for Artificial Intelligence Agriculture Disease Detection Under Cross-Cooperative Data Custody: A Computer Leakage-Audit Evaluation

Vertical federated learning (VFL) allows cooperatives that hold different feature blocks for the same farms to train a shared disease-detection model without pooling raw data. The privacy claim, however, is usually asserted rather than audited: a VFL protocol can be formally private under one leakage model and practically disclosive under another. The framework formalises the adversary's view as a transcript of exchanged gradients and intermediate representations, defines a leakage functional that separates feature-reconstruction risk from label-inference risk, and derives bounds on both under a declared Gaussian mechanism. A leakage-audit protocol is then specified around three quantities: reconstruction error of a held-out feature block, membership-inference advantage, and attribute-inference advantage, each with a stated decision threshold. Because no observational data were available, the framework is characterised analytically and through a replication-based design evaluation of the audit's operating characteristics under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64). The design evaluation indicates that 54 clusters are required for 80% power at a 5-percentage-point effect, and the audit is shown to be sensitive to the choice of leakage model rather than to model accuracy. The contribution is a reproducible audit specification that makes the privacy claim of agricultural VFL falsifiable before deployment.