An Energy-Driven Framework for Privacy-Aware Synthetic Data Generation
Abstract
Access to microdata in official statistics requires balancing disclosure risk and statistical utility. This paper proposes an interpretable, open-box energy-driven framework for privacy-aware synthesis of mixed-type tabular data, combining discriminative modelling, Bayesian-network proposals, Metropolis--Hastings sampling and post-generation optimization. Explicit energy components govern record-level plausibility, proximity to training records and diversity, while target selection balances fidelity and classification utility, making these trade-offs directly inspectable. Evaluation uses an artificially constructed individual-level dataset with demographic, behavioural and health-related variables. Validation combines marginal, bivariate and conditional fidelity diagnostics, diversity measures, nearest-neighbour analyses and membership inference attacks. An arrangement of classifiers trained on reference and synthetic data is evaluated on a common reference holdout. Repeated target selection assesses stability within a fixed candidate pool. Results show a convincing trade-off between fidelity and utility of the synthetic data generated by means of the proposed framework. Membership attacks perform close to chance across repeated splits, indicating limited training-membership discrimination under the evaluated protocol. The findings support the feasibility of balancing fidelity, utility and privacy-aware objectives through explicit generation and selection criteria, without implying formal privacy guarantees.
Related
- Phantoms and Disclosures: A Statistical Framework for Auditing Privacy in Synthetic Data
- Constraint-Embedded Fractal Self-Similar Generation of Synthetic Topologies for Receiving-End Power Grids with Rich Renewable Energy: A Hybrid Metaheuristic Approach
- A convolutional framework for detecting event-driven dynamics in energy price series
- Hankel-Koopman Finite-Horizon Energy Decomposition of Coupled Experimental Data: A Three-Phase Data-Driven Twin Forecasting Framework
- DagFlow AI: Agentic Synthetic Data Generation for Personalized Statistics Learning
BibTeX
@misc{massoli2026energydriven,
title = {{An Energy-Driven Framework for Privacy-Aware Synthetic Data Generation}},
author = {Pierpaolo Massoli and Fabio Spagnuolo},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2606.16488},
eprint = {2606.16488},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2606.16488},
}
RIS
TY - UNPB
TI - An Energy-Driven Framework for Privacy-Aware Synthetic Data Generation
AU - Pierpaolo Massoli
AU - Fabio Spagnuolo
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2606.16488
UR - https://arxiv.org/abs/2606.16488
AB - Access to microdata in official statistics requires balancing disclosure risk and statistical utility. This paper proposes an interpretable, open-box energy-driven framework for privacy-aware synthesis of mixed-type tabular data, combining discriminative modelling, Bayesian-network proposals, Metropolis--Hastings sampling and post-generation optimization. Explicit energy components govern record-level plausibility, proximity to training records and diversity, while target selection balances fidelity and classification utility, making these trade-offs directly inspectable. Evaluation uses an artificially constructed individual-level dataset with demographic, behavioural and health-related variables. Validation combines marginal, bivariate and conditional fidelity diagnostics, diversity measures, nearest-neighbour analyses and membership inference attacks. An arrangement of classifiers trained on reference and synthetic data is evaluated on a common reference holdout. Repeated target selection assesses stability within a fixed candidate pool. Results show a convincing trade-off between fidelity and utility of the synthetic data generated by means of the proposed framework. Membership attacks perform close to chance across repeated splits, indicating limited training-membership discrimination under the evaluated protocol. The findings support the feasibility of balancing fidelity, utility and privacy-aware objectives through explicit generation and selection criteria, without implying formal privacy guarantees.
ER -