QIF-DP: Privacy-Accounted, Leakage-Audited Budget Allocation for zCDP Query Answering

Abstract

In privacy-preserving statistical query systems, the released per-query privacy-budget vector can itself constitute privacy-relevant metadata. If budgets are computed from private workload statistics, the vector may reveal sparsity, imbalance, or dispersion even when the final answers are differentially private. We study this allocation-metadata problem under zero-concentrated differential privacy (zCDP). QIF-DP is a three-phase controller that releases zCDP diagnostics, computes the public budget vector by deterministic post-processing of those diagnostics and public confidence-interval targets, and answers the workload with Gaussian noise calibrated in zCDP units. The formal analysis and primary evaluation are confined to static, strictly disjoint scalar COUNT workloads within a central trusted-curator model. The utility contract emphasises confidence-interval target satisfaction and budget conservation rather than global MSE minimisation. Empirical results define a constrained operating region: in three hard-threshold zCDP regimes, QIF-DP improves target satisfaction over threshold-capped uniform allocation while reducing, but not eliminating, audit-detectable metadata dependence. The recommended deployment rule is therefore conditional: use QIF-DP only when pilot audits show target-satisfaction gain under acceptable residual leakage; otherwise use the threshold-capped uniform fallback.