Investigating User Perspectives on Differentially Private Text Privatization

Stephen Meisenbacher, Alexandra Klymenko, Alexander Karpp, Florian MatthesNAACL 2025

TL;DR

This paper explores how everyday users perceive text privatization using Differential Privacy (DP), a crucial step for real-world adoption. They found that users prioritize the quality and coherence of the privatized text, often choosing less privacy if the output is garbled. The mechanism used for privatization significantly impacts user choices, more so than sensitivity or data collection reasons.

What Problem Does This Paper Address?

Differentially Private Natural Language Processing (DP NLP) aims to protect sensitive information in text while preserving its meaning. A key challenge is "text privatization," where algorithms transform text to achieve this balance. While much research focuses on the technical aspects of DP NLP, there's a significant gap in understanding how users perceive these privacy-preserving mechanisms. This lack of user-centric understanding is a major hurdle for practical adoption. The authors address this by investigating what factors influence user preferences for text privatization, moving beyond abstract privacy budgets to concrete privatized text outputs.

Methods

The researchers employed a Factorial Survey Method (FSM), a robust technique for understanding how multiple factors influence user decisions. They designed a survey where 721 participants from around the globe were presented with vignettes. These vignettes described realistic data-sharing scenarios, each varying across four factors:

  • Mechanism Type: Four distinct DP text privatization methods were used (Truncated Exponential Mechanism (TEM), DP-MLM, DP-PROMPT, DP-BART), representing different approaches to text manipulation.
  • Privacy Budget (ε): Participants could adjust a slider representing the privacy level, from very low to very high privacy.
  • Reason for Data Collection: Scenarios varied whether data was collected for no reason, a vague reason (e.g., service improvement), or a specific reason (e.g., privacy protection).
  • Data Sensitivity: The text within the vignette was presented with varying levels of sensitivity (None, Low, High).

Participants then indicated their comfort level with sharing the text by selecting a privacy budget via the slider. Statistical analyses, including Chi-squared tests and Dunn's post-hoc tests, were used to determine the significance of each factor.

Datasets

The study did not use pre-existing datasets for training or evaluation. Instead, it generated its own data through a survey of 721 participants. The "datasets" in this context are the responses collected from these participants to the designed vignettes.

Key Insights and Ideas

  • Utility Trumps Privacy (When Utility is Poor): While users generally express a desire for strong privacy (as indicated by their IUIPC scores), their actual choices in the survey often favored less privacy if the privatized text was incoherent or grammatically incorrect. This highlights that for text privatization, "utility" (readability, coherence) is a prerequisite for users to even consider higher privacy levels.
  • Mechanism Matters Most: The type of DP mechanism used for privatization had the most significant impact on user choices. Generative approaches (DP-BART, DP-PROMPT) that produced more natural-sounding text led users to select higher privacy budgets, whereas non-generative methods (TEM, DP-MLM) often resulted in less private choices due to poorer output quality.
  • Context is Key, But Secondary: While scenario, data sensitivity, and the reason for data collection all played a role, they were less influential than the privatization mechanism itself. This suggests that while understanding the context of data sharing is important, the how of privatization is paramount for user acceptance.
  • "Acceptability Range" for Privatization: The study suggests an "acceptability range" for text privatization, roughly corresponding to 80-100% cosine similarity (slider 4 or 5). Outputs falling outside this range, particularly those with significant grammatical errors, were less preferred, even if they offered higher privacy.

Key Results

  • All four tested factors (scenario, sensitivity, mechanism, reason) were found to have a statistically significant impact on user choices for privacy budget.
  • The mechanism was the most salient factor, with generative mechanisms leading to higher privacy choices compared to non-generative ones.
  • Sensitivity also played a significant role, with higher sensitivity leading to a statistically significant correlation with lower privacy budgets (higher ɛ), though the effect size was weak.
  • Reason for data collection showed a significant relationship with privacy budget, but with a very small effect size. Dunn's post-hoc tests revealed a difference between "specific" and "none" reasons.
  • The study observed a skewed distribution of responses, with a strong preference for higher privacy slider values (closer to ε=∞), but this was heavily influenced by the quality of the privatized text.

Limitations and Future Work

Limitations:

  • Participant Understanding: The authors acknowledge that despite clear instructions, it's possible not all participants fully understood the technical nuances of DP text privatization.
  • Mechanism Output Quality: The privatized texts, especially from non-generative methods, sometimes contained obvious grammatical errors or non-ASCII characters, which could have biased user choices.
  • Sample Representativeness: The survey sample, while large, was not perfectly representative of the global population, with a disproportionately high number of respondents from South Africa. The study also did not analyze differences based on demographics like gender, profession, or education.
  • Domain Specificity: The study primarily focused on medical domain texts, which might not generalize to all text types.

Future Work:

  • Usability of DP NLP: Future research should focus on defining and achieving "usability" in DP text privatization, ensuring outputs are coherent, correct, and readable.
  • Contextualization: Further investigation into how different scenarios, data types, and purposes influence the required level of privacy is needed.
  • Human Studies: Continued user studies are crucial to align DP NLP research with real-world needs and perceptions, and to design methods for fostering acceptance of this technology.
  • Broader Samples and Domains: Replicating findings with more representative samples and diverse text domains is recommended.

Summary updated

BibTeX
@inproceedings{meisenbacher2025investigating,
  author = {Stephen Meisenbacher and Alexandra Klymenko and Alexander Karpp and Florian Matthes},
  title = {{Investigating User Perspectives on Differentially Private Text Privatization}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure