Collaborative Research: EAGER: Toward Provenance-Based Explainable Differential Privacy

This project supports research on provenance-based explanations for differentially private (DP) computations. DP provides rigorous protection for sensitive data, but systems often disclose only a privatized output and a declared privacy budget. Auditors, analysts, and data consumers may not know how an output was produced, why it is considered private, how privacy loss accumulated, or which execution-level choices influenced the release. The specific goal of this project is to develop a formal framework that treats carefully controlled mechanism metadata as part of the release itself, rather than as informal post-hoc annotation. Candidate provenance signals include sensitivity estimates, query composition, noise calibration, clipping behavior, adaptive thresholds, and budget-allocation decisions. For example, an explanation of a DP regional-income average may reveal subgroup sizes, query structure, or sensitivity bounds that influenced the result without exposing individual records. The project addresses the challenge that modern DP pipelines are often multi-stage and adaptive. Data-dependent computations, repeated releases, and correlated metadata can create privacy leakage even when an explanatory signal appears harmless in isolation. The research models the joint release of a DP output and selected provenance signals under a shared privacy budget. It characterizes which signals are public or derived by post-processing, which require privacy accounting, and how data-dependent signals should be perturbed, coarsened, selected, or suppressed. The project also studies provenance in learning-based systems, where the same signals may help models recover useful structure from noisy outputs but may also help learned adversaries sharpen inference about sensitive inputs. Attacker-aligned audits measures the additional inference risk from observing both the output and provenance. The resulting methods guides privacy- and learning-aware provenance-release policies that balance explanatory value, downstream utility, and end-to-end privacy risk. This project explores a fundamentally new direction in the data privacy space: treating explanation signals in DP systems as part of the release itself rather than post-hoc artifacts. Expected outcomes include formal joint-release foundations, prototype mechanisms, and evaluation methods that support transparent and accountable DP systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.