Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Measuring Privacy vs. Fidelity in Synthetic Social Media Datasets

Henry Tari, Adriana Iamnitchi

TL;DR

A methodology that quantifies privacy in synthetic text by framing re-identification as an authorship attribution attack is proposed and the privacy--fidelity tension in social media datasets is demonstrated.

Abstract

Synthetic data is increasingly used to support research without exposing sensitive user content. Social media data is one of the types of datasets that would hugely benefit from representative synthetic equivalents that can be used to bootstrap research and allow reproducibility through data sharing. However, recent studies show that (tabular) synthetic data is not inherently privacy-preserving. Much less is known, however, about the privacy risks of synthetically generated unstructured texts. This work evaluates the privacy of synthetic Instagram posts generated by three state-of-the-art large language models using two prompting strategies. We propose a methodology that quantifies privacy by framing re-identification as an authorship attribution attack. A RoBERTa-large classifier trained on real posts achieved 81\% accuracy in authorship attribution on real data, but only 16.5--29.7\% on synthetic posts, showing reduced, though non-negligible, risk. Fidelity was assessed via text traits, sentiment, topic overlap, and embedding similarity, confirming the expected trade-off: higher fidelity coincides with greater privacy leakage. This work provides a framework for evaluating privacy in synthetic text and demonstrates the privacy--fidelity tension in social media datasets.

Measuring Privacy vs. Fidelity in Synthetic Social Media Datasets

TL;DR

A methodology that quantifies privacy in synthetic text by framing re-identification as an authorship attribution attack is proposed and the privacy--fidelity tension in social media datasets is demonstrated.

Abstract

Synthetic data is increasingly used to support research without exposing sensitive user content. Social media data is one of the types of datasets that would hugely benefit from representative synthetic equivalents that can be used to bootstrap research and allow reproducibility through data sharing. However, recent studies show that (tabular) synthetic data is not inherently privacy-preserving. Much less is known, however, about the privacy risks of synthetically generated unstructured texts. This work evaluates the privacy of synthetic Instagram posts generated by three state-of-the-art large language models using two prompting strategies. We propose a methodology that quantifies privacy by framing re-identification as an authorship attribution attack. A RoBERTa-large classifier trained on real posts achieved 81\% accuracy in authorship attribution on real data, but only 16.5--29.7\% on synthetic posts, showing reduced, though non-negligible, risk. Fidelity was assessed via text traits, sentiment, topic overlap, and embedding similarity, confirming the expected trade-off: higher fidelity coincides with greater privacy leakage. This work provides a framework for evaluating privacy in synthetic text and demonstrates the privacy--fidelity tension in social media datasets.
Paper Structure (23 sections, 4 equations, 5 figures, 8 tables)

This paper contains 23 sections, 4 equations, 5 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset
  4. Generating Synthetic Posts using LLMs
  5. Identifying Representative Sample Size for Data Generation
  6. Prompting Strategies
  7. The Privacy Risks of Social Media Data
  8. Re-Identification Attack
  9. Stylometric Methods
  10. N-gram Methods
  11. TF-IDF Features
  12. Transformer-Based Methods
  13. Results of Re-Identification Attacks
  14. Single-Model Performance
  15. Combined Models
  16. ...and 8 more sections

Figures (5)

  • Figure 1: Distribution of number of posts per writer in the dataset. There are 30 bins with the size of 156.7. While the majority of authors have fewer than 1000 posts, some have over 4000, with a maximum of 6018.
  • Figure 2: Distribution of post lengths. Most posts are shorter than 100 words with the median text length being 14.
  • Figure 3: Sentiment distribution for all LLMs compared to real data under example-based and persona-based prompting.
  • Figure 4: Topic overlap between real and DeepSeek-generated data. Shared topics remain stable across prompts, but persona-based prompting produces many more unique topics.
  • Figure 5: t-SNE visualization of generated posts by models per prompting strategy and original posts.