[Paper Review] Using Neural Generative Models to Release Synthetic Twitter Corpora with Reduced Stylometric Identifiability of Users

Alexander G. Ororbia, Fridolin Linder+1 more|arXiv (Cornell University)|Jun 3, 2016

Authorship Attribution and Profiling3 citations

TL;DR

This paper proposes a neural generative model that synthesizes Twitter corpora to reduce stylometric re-identification risk while preserving user-specific word distributions. By conditioning on individual users and tuning the temperature hyperparameter, the method achieves a superior risk-utility trade-off compared to redaction and iterative translation, significantly lowering re-identification risk with minimal utility loss.

ABSTRACT

We present a method for generating synthetic versions of Twitter data using neural generative models. The goal is protecting individuals in the source data from stylometric re-identification attacks while still releasing data that carries research value. Specifically, we generate tweet corpora that maintain user-level word distributions by augmenting the neural language models with user-specific components. We compare our approach to two standard text data protection methods: redaction and iterative translation. We evaluate the three methods on measures of risk and utility. We define risk following the stylometric models of re-identification, and we define utility based on two general word distribution measures and two common text analysis research tasks. We find that neural models are able to significantly lower risk over previous methods with little cost to utility. We also demonstrate that the neural models allow data providers to actively control the risk-utility trade-off through model tuning parameters. This work presents promising results for a new tool addressing the problem of privacy for free text and sharing social media data in a way that respects privacy and is ethically responsible.

Motivation & Objective

• To address the ethical and privacy challenges in releasing social media data, especially after users delete content or request to be forgotten.
• To reduce the risk of stylometric re-identification of users in shared Twitter corpora without sacrificing analytical utility.
• To develop a flexible, data-driven method that allows privacy tuning through model hyperparameters, unlike blunt techniques like redaction or translation.
• To evaluate the method empirically against standard privacy protection techniques using both risk and utility metrics.

Proposed method

• A novel neural architecture, the Delta-RNN, is trained to generate synthetic tweets conditioned on individual user language features, preserving user-specific word distributions.
• The model uses a temperature parameter to control the noise level in the output distribution, enabling direct tuning of the risk-utility trade-off.
• User-specific conditioning is implemented by incorporating user embeddings into the recurrent network architecture, allowing personalized generation.
• The method generates character-level sequences using a language model trained on user-specific tweet corpora, ensuring linguistic fidelity.
• The approach is compared to redaction and iterative translation using stylometric re-identification risk and word distribution similarity as key evaluation metrics.
• The temperature hyperparameter is systematically varied to assess its impact on both risk (re-identification accuracy) and utility (distributional similarity and task performance).

Experimental results

Research questions

• RQ1 Can neural generative models effectively reduce stylometric re-identification risk in synthetic Twitter corpora while maintaining user-level linguistic patterns?
• RQ2 How does the risk-utility trade-off of neural generation compare to traditional methods like redaction and iterative translation?
• RQ3 To what extent can the temperature hyperparameter be used to actively tune the balance between privacy and data utility?
• RQ4 Does the neural model preserve the distributional and task-specific utility of the original data across common NLP tasks?
• RQ5 How robust is the proposed method against multiple stylometric attack models and feature sets?

Key findings

• The neural model significantly reduces re-identification risk—achieving less than 10% risk at optimal temperature settings—compared to over 50% risk with redaction.
• The Delta-RNN model achieves the best risk-utility trade-off, outperforming both redaction and translation across all evaluation metrics.
• Lower temperature values reduce risk by suppressing rare, identifying words, but also reduce utility due to decreased distributional similarity.
• The temperature parameter enables precise control over the risk-utility trade-off, offering a tunable privacy knob not available in redaction or translation.
• For sentiment classification and n-gram similarity, the neural model maintains high utility even at low temperatures, with minimal performance degradation.
• The model’s performance is robust across multiple stylometric attack models, validating its effectiveness under diverse re-identification scenarios.