[Paper Review] HARMONIC: Harnessing LLMs for Tabular Data Synthesis and Privacy Protection

Yuxin Wang, Duanyu Feng+6 more|arXiv (Cornell University)|Aug 6, 2024

Advanced Data Storage Technologies4 citations

TL;DR

HARMONIC proposes a novel framework that leverages instruction-fine-tuned large language models (LLMs) to generate high-quality, privacy-preserving synthetic tabular data by modeling inter-row relationships via a k-nearest neighbors-inspired prompting strategy. The method achieves performance comparable to real data in downstream LLM tasks while significantly reducing privacy leakage, as validated by new metrics DLT (data leakage test) and LLE (LLM efficiency).

ABSTRACT

Data serves as the fundamental foundation for advancing deep learning, particularly tabular data presented in a structured format, which is highly conducive to modeling. However, even in the era of LLM, obtaining tabular data from sensitive domains remains a challenge due to privacy or copyright concerns. Hence, exploring how to effectively use models like LLMs to generate realistic and privacy-preserving synthetic tabular data is urgent. In this paper, we take a step forward to explore LLMs for tabular data synthesis and privacy protection, by introducing a new framework HARMONIC for tabular data generation and evaluation. In the tabular data generation of our framework, unlike previous small-scale LLM-based methods that rely on continued pre-training, we explore the larger-scale LLMs with fine-tuning to generate tabular data and enhance privacy. Based on idea of the k-nearest neighbors algorithm, an instruction fine-tuning dataset is constructed to inspire LLMs to discover inter-row relationships. Then, with fine-tuning, LLMs are trained to remember the format and connections of the data rather than the data itself, which reduces the risk of privacy leakage. In the evaluation part of our framework, we develop specific privacy risk metrics DLT for LLM synthetic data generation, as well as performance evaluation metrics LLE for downstream LLM tasks. Our experiments find that this tabular data generation framework achieves equivalent performance to existing methods with better privacy, which also demonstrates our evaluation framework for the effectiveness of synthetic data and privacy risks in LLM scenarios.

Motivation & Objective

• Address the challenge of generating realistic tabular data from sensitive domains due to privacy and regulatory constraints.
• Overcome the privacy risks inherent in LLM-based data synthesis, where models may memorize and leak training data.
• Develop a framework that enables LLMs to generate synthetic tabular data without memorizing real data instances.
• Introduce new evaluation metrics tailored for assessing both downstream LLM task performance and privacy risks in synthetic data.
• Demonstrate that LLM-based synthetic data can be both effective and privacy-safe when properly fine-tuned and evaluated.

Proposed method

• Construct an instruction fine-tuning dataset inspired by k-nearest neighbors (kNN) to guide LLMs in learning inter-row relationships in tabular data.
• Fine-tune large-scale LLMs on the kNN-based instruction dataset to memorize data structure and relationships rather than raw data values.
• Use the fine-tuned LLM to generate synthetic tabular data by autoregressively predicting tokens while preserving feature-level and row-level dependencies.
• Propose DLT (Data Leakage Test) as a novel metric to quantify privacy risk by measuring the likelihood of reconstructing real data from synthetic outputs.
• Introduce LLE (LLM Efficiency) as a metric to evaluate the effectiveness of synthetic data in downstream LLM tasks, such as fine-tuning and inference.
• Apply the framework to four real-world tabular datasets for classification, comparing performance and privacy across multiple baselines.

Experimental results

Research questions

• RQ1 Can instruction-fine-tuned LLMs generate synthetic tabular data that preserves data structure and relationships without memorizing real instances?
• RQ2 How does the privacy risk of LLM-generated synthetic data compare to existing methods, especially under reconstruction attacks?
• RQ3 To what extent can synthetic data generated by LLMs perform in downstream LLM tasks compared to real data?
• RQ4 Can the proposed evaluation metrics DLT and LLE effectively quantify privacy and utility in LLM-based synthetic tabular data?
• RQ5 Does the kNN-inspired fine-tuning strategy improve the consistency and reliability of synthetic data generation compared to baseline LLM methods?

Key findings

• HARMONIC-generated synthetic data achieves comparable performance to real data in downstream LLM tasks, with an average drop of less than 5% across four datasets.
• On the DI dataset, HARMONIC outperforms the real training set in LLM efficiency (LLE), indicating strong downstream utility.
• HARMONIC achieves the best or second-best scores in all three privacy metrics (NRS, DCR, DLT) across all datasets, with DLT scores indicating significantly lower data leakage risk.
• The method exhibits superior stability, with lower standard deviation in performance across runs compared to GReaT and RTF, indicating consistent and reliable generation.
• Traditional synthetic data methods like SMOTE and TVAE perform poorly in downstream LLM tasks, suggesting they are unsuitable for LLM fine-tuning.
• The framework’s design inherently resists reconstruction attacks, as an attacker would need nearly the entire k-neighbor set (k=5) to reconstruct a single record, unlike methods like GReaT that are vulnerable to partial record reconstruction.