Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Initial Exploration of Zero-Shot Privacy Utility Tradeoffs in Tabular Data Using GPT-4

Bishwas Mandal, George Amariucai, Shuangqing Wei

TL;DR

This work explores privacy-utility tradeoffs in tabular data using GPT-4 as a zero-shot sanitization tool. By converting tabular rows to text and appending sanitization instructions in prompts, the authors create a sanitizing function via$f(.)=g(p(.))$, yielding sanitized data $\,\hat{D}$ that preserves utility features $D_U$ while obscuring private features $D_P$. Across the UCI Adult dataset with two task configurations, GPT-4-based prompts achieve privacy protection comparable to adversarial baselines and, in some configurations, even superior utility retention, though fairness metrics are not uniformly satisfied. The study highlights the potential of LLMs for privacy-preserving data sharing without additional training, while acknowledging limitations in fairness and pointing to future improvements with advancing models.

Abstract

We investigate the application of large language models (LLMs), specifically GPT-4, to scenarios involving the tradeoff between privacy and utility in tabular data. Our approach entails prompting GPT-4 by transforming tabular data points into textual format, followed by the inclusion of precise sanitization instructions in a zero-shot manner. The primary objective is to sanitize the tabular data in such a way that it hinders existing machine learning models from accurately inferring private features while allowing models to accurately infer utility-related attributes. We explore various sanitization instructions. Notably, we discover that this relatively simple approach yields performance comparable to more complex adversarial optimization methods used for managing privacy-utility tradeoffs. Furthermore, while the prompts successfully obscure private features from the detection capabilities of existing machine learning models, we observe that this obscuration alone does not necessarily meet a range of fairness metrics. Nevertheless, our research indicates the potential effectiveness of LLMs in adhering to these fairness metrics, with some of our experimental results aligning with those achieved by well-established adversarial optimization techniques.

Initial Exploration of Zero-Shot Privacy Utility Tradeoffs in Tabular Data Using GPT-4

TL;DR

This work explores privacy-utility tradeoffs in tabular data using GPT-4 as a zero-shot sanitization tool. By converting tabular rows to text and appending sanitization instructions in prompts, the authors create a sanitizing function via, yielding sanitized data that preserves utility features while obscuring private features . Across the UCI Adult dataset with two task configurations, GPT-4-based prompts achieve privacy protection comparable to adversarial baselines and, in some configurations, even superior utility retention, though fairness metrics are not uniformly satisfied. The study highlights the potential of LLMs for privacy-preserving data sharing without additional training, while acknowledging limitations in fairness and pointing to future improvements with advancing models.

Abstract

We investigate the application of large language models (LLMs), specifically GPT-4, to scenarios involving the tradeoff between privacy and utility in tabular data. Our approach entails prompting GPT-4 by transforming tabular data points into textual format, followed by the inclusion of precise sanitization instructions in a zero-shot manner. The primary objective is to sanitize the tabular data in such a way that it hinders existing machine learning models from accurately inferring private features while allowing models to accurately infer utility-related attributes. We explore various sanitization instructions. Notably, we discover that this relatively simple approach yields performance comparable to more complex adversarial optimization methods used for managing privacy-utility tradeoffs. Furthermore, while the prompts successfully obscure private features from the detection capabilities of existing machine learning models, we observe that this obscuration alone does not necessarily meet a range of fairness metrics. Nevertheless, our research indicates the potential effectiveness of LLMs in adhering to these fairness metrics, with some of our experimental results aligning with those achieved by well-established adversarial optimization techniques.
Paper Structure (11 sections, 2 equations, 7 figures, 3 tables, 1 algorithm)

This paper contains 11 sections, 2 equations, 7 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Formulation and Methodology
  4. Assumptions and Threat Model
  5. Existing Sanitization Mechanisms (Baselines)
  6. Experimental Setup
  7. Dataset
  8. Performance Evaluation Metrics
  9. Implementation Details
  10. Result and Discussions
  11. Conclusion

Figures (7)

  • Figure 1: Overview of the proposed data sanitization technique using GPT-4. Our approach involves initializing the GPT-4 model with a prompt that includes the conversion of tabular data points into textual format, followed by the incorporation of sanitization instructions. Additionally, we incorporate formatting instructions into the prompt to ensure that the generated outputs are suitable for programmatic utilization. The presented prompt above pertains to the prompt designated for Task 1.
  • Figure 2: ALFR and UAE-PUPET architecture
  • Figure 3: GPT-4 zero-shot classification prompt, GPT-4 (C)
  • Figure 4: Fairness metrics for Task 1 and Task 2. The X-axis represents different ML models, and the Y-axis represents the fairness scores for those ML models, without sanitization (No PM), and after using different sanitization techniques.
  • Figure 5: Visualization of distortion (noise) applied to continuous variables by different sanitization mechanisms for Task 1 and Task 2. Upper row shows distortion applied for Task 1 and lower row shows distortion applied for Task 2.
  • ...and 2 more figures