[Paper Review] Auditing and Generating Synthetic Data with Controllable Trust Trade-offs

Brian Belgodere, Pierre Dognin+12 more|arXiv (Cornell University)|Apr 21, 2023

Privacy-Preserving Technologies in Data4 citations

TL;DR

This paper proposes a holistic auditing framework for synthetic data that evaluates trustworthiness across fidelity, utility, privacy, fairness, and robustness, using a controllable trustworthiness index to guide model selection. It introduces TrustFormers—differentiable generative models trained via trustworthiness-driven cross-validation—demonstrating superior performance across diverse datasets (e.g., law school, healthcare) under varying trade-offs, with synthetic data ranking highest in 6 of 10 weighted evaluations when uncertainty is accounted for.

ABSTRACT

Real-world data often exhibits bias, imbalance, and privacy risks. Synthetic datasets have emerged to address these issues. This paradigm relies on generative AI models to generate unbiased, privacy-preserving data while maintaining fidelity to the original data. However, assessing the trustworthiness of synthetic datasets and models is a critical challenge. We introduce a holistic auditing framework that comprehensively evaluates synthetic datasets and AI models. It focuses on preventing bias and discrimination, ensures fidelity to the source data, assesses utility, robustness, and privacy preservation. We demonstrate the framework's effectiveness by auditing various generative models across diverse use cases like education, healthcare, banking, and human resources, spanning different data modalities such as tabular, time-series, vision, and natural language. This holistic assessment is essential for compliance with regulatory safeguards. We introduce a trustworthiness index to rank synthetic datasets based on their safeguards trade-offs. Furthermore, we present a trustworthiness-driven model selection and cross-validation process during training, exemplified with "TrustFormers" across various data types. This approach allows for controllable trustworthiness trade-offs in synthetic data creation. Our auditing framework fosters collaboration among stakeholders, including data scientists, governance experts, internal reviewers, external certifiers, and regulators. This transparent reporting should become a standard practice to prevent bias, discrimination, and privacy violations, ensuring compliance with policies and providing accountability, safety, and performance guarantees.

Motivation & Objective

• To address the lack of comprehensive, multi-dimensional auditing for synthetic data across diverse modalities and regulatory requirements.
• To develop a unified trustworthiness index that quantifies trade-offs between key safeguards like privacy, fairness, and utility.
• To enable controllable trustworthiness in synthetic data generation through model selection and cross-validation guided by the trustworthiness index.
• To ensure compliance with evolving AI regulations such as the EU AI Act and U.S. Algorithmic Accountability Act by providing transparent, auditable synthetic data.
• To foster collaboration among stakeholders—including regulators, data scientists, and governance teams—through standardized reporting and risk transparency.

Proposed method

• Proposes a holistic auditing framework evaluating synthetic data across five trust pillars: fidelity, utility, privacy, fairness, and robustness.
• Introduces a trustworthiness index that aggregates weighted scores across trust dimensions, enabling trade-off analysis.
• Employs a trustworthiness-driven cross-validation process during training to select models that meet desired safeguard trade-offs.
• Develops TrustFormers—differentiable generative models trained with loss functions optimized for specific trustworthiness index configurations.
• Applies uncertainty-aware evaluation using $ R^{ au}_{ au} $ to rank synthetic data under statistical volatility, improving reliability of rankings.
• Uses real-world datasets across tabular, time-series, vision, and NLP modalities to validate the framework across education, healthcare, banking, and HR use cases.

Experimental results

Research questions

• RQ1 How can synthetic data be holistically audited across multiple trust dimensions—fidelity, utility, privacy, fairness, and robustness—simultaneously?
• RQ2 What is an effective way to quantify and control trade-offs between competing trust objectives in synthetic data generation?
• RQ3 Can a trustworthiness index improve model selection and cross-validation to produce more reliable and compliant synthetic data?
• RQ4 How does accounting for uncertainty in data splits affect the ranking and reliability of synthetic data performance?
• RQ5 To what extent can TrustFormers outperform existing baselines (e.g., SDV, DP-GAN, PATE-GAN) under controlled trustworthiness trade-offs?

Key findings

• TrustFormers trained with the trustworthiness index outperformed other synthetic data in 6 out of 10 weighting configurations on the law school dataset, with the highest mean trustworthiness index.
• When uncertainty is accounted for using $ R^{ au}_{ au} $, TrustFormer synthetic data reclaims the top rank in the leaderboards, demonstrating robustness to data split variability.
• The private TrustFormer with $ ho = 3 $ and $ ho = 1 $ achieved the highest trustworthiness index scores (39 and 38 respectively) across multiple metrics, outperforming even non-private baselines like SDV-CTGAN.
• The framework successfully identified that models like DP-GAN with $ ho = 3 $ achieved near-perfect privacy but suffered in utility and fairness, highlighting trade-offs.
• The trustworthiness index effectively ranked synthetic data across diverse modalities and use cases, with TrustFormers consistently ranking in the top 10% across all evaluations.
• The inclusion of uncertainty-aware evaluation reduced volatility in rankings, showing that TrustFormers maintain high performance even under statistical fluctuations in data splits.