D3.8: Privacy-preserving analytics: first release

This deliverable presents the first release of the privacy-preserving analytics workflow developed within the HEREDITARY project. It builds on the Hereditary Data Network (HDN) and addresses the need to perform analytics over sensitive medical data while preserving institutional control and reducing disclosure risks. The deliverable revises the privacy setting of the HDN by introducing PRISM, a query-level privacy-risk assessment framework designed in collaboration with WP7 and grounded in the legal and ethical requirements for federated processing of sensitive health data. PRISM enables calibrated and explainable assessment based on the query, ontology annotations, and, where required, data-level statistics.

The deliverable also positions Privacy-Preserving Synthetic Data Generation (PP-SDG) as a controlled fallback mechanism for queries and analytics tasks that cannot be safely answered over real data. We focus on tabular synthesis methods and survey privacy metrics for estimating risks associated to synthetic data releases. We organise these metrics in a taxonomy based on their computational structure and threat model, and we operationalise them in PrivEval, an interactive framework supporting data owners in assessing residual disclosure risk and metric applicability.

The experimental evaluation focuses on the tabular ALS data, at the core of many analyses of the clinical partners. Among the PP-SDG methods considered (DPGAN, PATE-GAN, and PrivBayes), PrivBayes provides the strongest observed utility, preserving selected marginal distributions and achieving higher utility than DPGAN and PATE-GAN. At the same time, the empirical privacy metrics do not indicate a higher residual disclosure risk for PrivBayes, suggesting that PrivBayes can achieve Differential Privacy (DP) without significantly degrading the ALS data. At the same time, we identified limitations of current PP-SDG methods, including scalability constraints, weakened dependencies between attributes, and possible generation of clinically implausible records.

Finally, the deliverable evaluates k-means as a representative downstream federated analytics task. The results show that federated k-means over real ALS data approximates the corresponding centralised clustering with limited distortion, while the main loss of utility is introduced by the synthetic data generation step. The deliverable also introduces a SPARQL-based implementation path for federated k-means, aimed at closer integration with the HDN execution model. Together, these results establish the foundations of the HEREDITARY privacy-preserving analytics pipeline: PRISM provides the governance layer, PP-SDG provides a candidate fallback mechanism, PrivEval provides the empirical assessment layer, and the ALS experiments identify the main limitations to address in subsequent iterations.