[论文解读] Beyond Privacy: Navigating the Opportunities and Challenges of Synthetic Data

Boris van Breugel, Mihaela van der Schaar|arXiv (Cornell University)|Apr 7, 2023

Privacy-Preserving Technologies in Data被引用 14

一句话总结

This perspective surveys how data-driven synthetic data can go beyond privacy to enable augmentation, domain adaptation, simulation, fairness, and user-prompted data, while highlighting fundamental challenges in trust, metrics, and applicability.

ABSTRACT

Generating synthetic data through generative models is gaining interest in the ML community and beyond. In the past, synthetic data was often regarded as a means to private data release, but a surge of recent papers explore how its potential reaches much further than this -- from creating more fair data to data augmentation, and from simulation to text generated by ChatGPT. In this perspective we explore whether, and how, synthetic data may become a dominant force in the machine learning world, promising a future where datasets can be tailored to individual needs. Just as importantly, we discuss which fundamental challenges the community needs to overcome for wider relevance and application of synthetic data -- the most important of which is quantifying how much we can trust any finding or prediction drawn from synthetic data.

研究动机与目标

• Motivate synthetic data as a versatile alternative to real data beyond privacy constraints.
• Define data-driven synthetic data and its potential to tailor datasets.
• Systematically review major use cases and their opportunities and challenges.
• Identify general challenges and propose directions for metrics, evaluation, and trustworthiness.
• Outline a roadmap for broader adoption through standardized practices and benchmarks.

提出的方法

• Define data-driven synthetic data and distinguish it from hand-crafted synthetic data.
• Survey use cases: privacy, augmentation, domain adaptation, data-driven simulations, fairness, and user-prompted data.
• Discuss challenges and trade-offs for each use case (privacy-utility, realism, representativeness, etc.).
• Highlight general challenges and open questions in metrics, model choice, outliers, downstream impact, and verification.
• Propose criteria for trustworthy evaluation and data verification mechanisms.

实验结果

研究问题

• RQ1 What opportunities do synthetic data offer beyond privacy, and what applications look most promising?
• RQ2 What are the core challenges in trusting and evaluating synthetic data, and how might metrics and benchmarks address them?
• RQ3 How can synthetic data be effectively used across augmentation, domain adaptation, and simulation, while managing fairness and privacy concerns?
• RQ4 What guidance is needed to select models, standards, and verification procedures to enable broader adoption?

主要发现

• Synthetic data has potential to replace or augment real data, enabling privacy-preserving, fairer, more robust, and customizable datasets.
• There is a privacy-utility trade-off in synthetic data generation, with no perfect privacy metric and challenges in future-proof privacy guarantees.
• Domain adaptation, augmentation, and data-driven simulation can improve data efficiency and model reliability, especially in underrepresented settings.
• Fairness with synthetic data is feasible but may incur utility losses and requires careful alignment with downstream deployment contexts.
• User-prompted synthetic data (e.g., ChatGPT-like outputs) demonstrates broad applications but raises trust, copyright, and authenticity concerns that demand urgent solutions.
• The field faces fundamental open questions in applicability, quality measurement, and verification that hinder widespread adoption.