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
In the social sciences, Twitter serves as a vital resource for understanding human behavior. However, sharing these datasets is fraught with legal and ethical challenges. Beyond the "right to be forgotten," which requires researchers to delete data when users delete tweets, there is the persistent threat of stylometric re-identification. Even anonymized text contains "fingerprints"—unique patterns in syntax and word choice—that allow attackers to link data back to specific individuals. This research addresses the critical "Risk-Utility Trade-off": how to release data that remains analytically useful while ensuring it is impossible to link back to a real person.
Methodology: Multi-User Conditional Synthesis
The authors propose a generative approach called Multi-User Conditional Synthesis. Instead of sharing original tweets, they train neural models to learn the underlying statistical distribution of the dataset and generate synthetic versions.
As illustrated in the architecture above, the researchers modified standard Recurrent Neural Networks (RNNs)—including Elman RNNs, GRUs, and a novel "Delta-RNN"—to accept a User ID as an additional input. This allows a single model to capture the general rules of language while simultaneously modeling the specific stylistic quirks of hundreds of different users.
A key feature of this system is the temperature parameter ($\tau$), which acts as a "privacy knob." By adjusting the temperature during text generation, data providers can control the level of stochasticity: lower temperatures produce text very similar to the original (high utility, higher risk), while higher temperatures introduce more noise to thwart identification (lower risk, lower utility).
Evaluating Risk and Utility
To evaluate the effectiveness of their approach, the authors simulated a stylometric attack using a Support Vector Machine (SVM) classifier. The attacker's goal is to identify the true author of a tweet based on a variety of linguistic features.
The features used for these attacks (detailed in the table above) range from simple uni-grams to complex syntactic category pairs. The study utilized a corpus of approximately 448,000 tweets from 627 users to test how well synthetic data could resist these sophisticated identification techniques.
Key Results and Insights
The study yielded several significant findings:
- Superior Risk Reduction: While original data allowed for 82% re-identification accuracy, the neural models slashed this risk. The Delta-RNN, at moderate temperatures, reduced risk to 44%, and at high temperatures, it dropped as low as 8%.
- High Utility Preservation: Despite the privacy gains, the synthetic data remained highly representative. In sentiment analysis tasks, synthetic data achieved a cosine similarity of 0.82 to 0.93 compared to the original data.
- The Efficiency Frontier: Neural models outperformed traditional baselines like Redaction (removing hashtags/handles) and Iterative Translation (back-translation through multiple languages). They provided a better balance, maintaining higher analytical utility for any given level of risk reduction.
Conclusion
Synthetic data generation via neural models represents a paradigm shift in data privacy for natural language. By shifting the focus from the original words to the statistical properties of the data, researchers can share datasets that are ethically responsible and legally compliant. Future work may explore more complex architectures like Variational Autoencoders (VAEs) to further refine the modeling of complex user features at scale.
Summary updated