On the Vulnerability of Text Sanitization

TL;DR

This paper critically evaluates the privacy protection effectiveness of text sanitization by developing theoretically optimal and practical reconstruction attacks, revealing significant vulnerabilities and prompting a reassessment of current sanitization methods.

Contribution

It introduces theoretically grounded and practical reconstruction attacks that outperform existing methods, providing a more accurate assessment of text sanitization privacy risks.

Findings

One attack improved attack success rate by 46.4% over baseline.

Revealed significant vulnerabilities in current text sanitization methods.

Provided bounds on attack success rate as benchmarks for evaluation.

Abstract

Text sanitization, which employs differential privacy to replace sensitive tokens with new ones, represents a significant technique for privacy protection. Typically, its performance in preserving privacy is evaluated by measuring the attack success rate (ASR) of reconstruction attacks, where attackers attempt to recover the original tokens from the sanitized ones. However, current reconstruction attacks on text sanitization are developed empirically, making it challenging to accurately assess the effectiveness of sanitization. In this paper, we aim to provide a more accurate evaluation of sanitization effectiveness. Inspired by the works of Palamidessi et al., we implement theoretically optimal reconstruction attacks targeting text sanitization. We derive their bounds on ASR as benchmarks for evaluating sanitization performance. For real-world applications, we propose two practical reconstruction attacks based on these theoretical findings. Our experimental results underscore the necessity of reassessing these overlooked risks. Notably, one of our attacks achieves a 46.4% improvement in ASR over the state-of-the-art baseline, with a privacy budget of epsilon=4.0 on the SST-2 dataset. Our code is available at: https://github.com/mengtong0110/On-the-Vulnerability-of-Text-Sanitization.