Measuring Privacy vs. Fidelity in Synthetic Social Media Datasets

2026-03-04

AI Summary

This study investigates the trade-off between privacy preservation and data fidelity in synthetic social media text generated by large language models, with a focus on re-identification risks in unstructured text. It introduces authorship attribution attacks to this domain and establishes a systematic evaluation framework: using a RoBERTa-large classifier to perform attacks on synthetic Instagram posts generated by three state-of-the-art large language models under two prompting strategies, while assessing fidelity through metrics including textual features, sentiment, topic overlap, and embedding similarity. Experimental results show that authorship attribution accuracy drops from 81% on real data to 16.5%–29.7% on synthetic data, confirming a substantial yet non-negligible reduction in privacy risk. Moreover, the study reveals a positive correlation between fidelity and privacy leakage, highlighting an inherent tension between the two objectives.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Privacy
• Computer Vision: Bias, Fairness & Privacy

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

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.

Problem

• synthetic data
• privacy
• fidelity
• social media
• authorship attribution

Innovation

• synthetic text
• privacy evaluation
• authorship attribution
• fidelity-privacy trade-off
• large language models