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.
Introduction: The Privacy-Utility Trade-off in Social Media Data
Sharing social media data for research is a complex challenge. While social scientists require large datasets to understand human behavior, users have a right to privacy and a "right to be forgotten." When researchers release "anonymized" text data, they typically strip away usernames. However, "stylometry"—the study of linguistic style—can often link an anonymous text back to its author based on unique patterns in word choice, punctuation, and grammar. On Twitter, this is a major privacy risk; even if a user deletes a tweet, a researcher’s archived copy might still be used to identify them.
Existing protection methods are often "blunt instruments." Redaction (removing hashtags and handles) is easily bypassed, and iterative translation (e.g., translating English to Arabic and back) often mangles the data's meaning, destroying its "utility" for researchers. This paper introduces a way to generate synthetic tweet corpora that mimic a user's style without exposing their identity using neural generative models.
Multi-User Conditional Synthesis
The authors propose a Multi-User Conditional Synthesis Model. Instead of training a separate model for every person, they built a single large neural network that takes a "User ID" as an additional input. This allows the model to learn general English grammar from the whole group while still capturing individual-specific nuances.
Key technical components include:
- Architectures: Three types of Recurrent Neural Networks (RNNs) were tested: the standard Elman RNN, the Gated Recurrent Unit (GRU), and the Delta-RNN (designed to better capture long-term dependencies).
- Character-Level Modeling: Because tweets are full of typos, emojis, and slang, the models predict text one character at a time rather than word by word.
- Temperature Tuning: A "knob" called temperature ($\tau$) was introduced. A low temperature makes the model pick the most likely characters (high utility, higher risk), while a high temperature adds randomness, making the output harder to trace (lower utility, lower risk).
Evaluating Risk and Utility
The study used a dataset of tweets from 627 German Twitter users (English-language posts). To evaluate the effectiveness of the synthetic data, the researchers simulated a "stylometric attack" where an adversary tries to match the synthetic tweets in a release set back to the real users in an attack set.
The attack model utilized a variety of features, as detailed in Figure 2, to capture the unique "fingerprint" of each user. On the original, unaltered data, a simple machine learning classifier could correctly identify the true author out of 627 candidates 82% of the time.
Key Results: The Pareto Frontier of Privacy
Neural models provide a significantly better trade-off than traditional methods. For many tasks, the neural models could cut the re-identification risk in half while keeping the data almost as useful as the original.
- Risk Reduction: The Elman RNN model at a specific temperature reduced the top-1 re-identification risk from 82% down to just 10%.
- Utility Preservation: Despite the massive drop in risk, the synthetic data maintained high utility. In a sentiment analysis task, the synthetic data achieved a cosine similarity of 0.80 to 0.93 compared to the original data.
- Failure of Redaction: Simply removing hashtags and handles only dropped the re-identification risk to 52%, which remains unacceptably high for many privacy standards.
- Delta-RNN Performance: The Delta-RNN architecture generally offered the best balance, providing the highest utility for a given level of privacy risk.
Conclusion
This work presents promising results for a new tool addressing the problem of privacy for free text. By using neural synthesis, data providers can actively control the risk-utility trade-off through model tuning parameters, sharing social media data in a way that respects privacy and is ethically responsible. Future work aims to explore Variational Autoencoders (VAEs) to generate even richer, more diverse synthetic text.
Summary updated