Abstract
In this paper we consider methods for sharing free text Twitter data, with the goal of protecting the privacy of individuals in the data while still releasing data that carries research value, i.e. minimizes risk and maximizes utility. We propose three protection methods: simple redaction of hashtags and twitter handles, an epsilon-differentially private Multinomial-Dirichlet synthesizer, and novel synthesis models based on a neural generative model. We evaluate these three methods using empirical measures of risk and utility. We define risk based on possible identification of users in the Twitter data, and we define utility based on two general language measures and two model-based tasks. We find that redaction maintains high utility for simple tasks but at the cost of high risk, while some neural synthesis models are able to produce higher levels of utility, even for more complicated tasks, while maintaining lower levels of risk. In practice, utility and risk present a trade-off, with some methods offering lower risk or higher utility. This work presents possible methods to approach the problem of privacy for free text and which methods could be used to meet different utility and risk thresholds.
Introduction
In the social sciences, sharing datasets is crucial for replicable research. However, Twitter data presents a unique privacy minefield. Users have a "right to be forgotten" (deleting tweets), and even when usernames are removed, "stylometry"—the unique linguistic fingerprint of a writer—can be used to re-identify anonymous authors with high accuracy.
The challenge is the classic Risk-Utility Trade-off:
- Risk: The probability that an attacker can link an "anonymized" tweet back to a real person.
- Utility: The degree to which the released data remains useful for research tasks like sentiment analysis or trend prediction.
The Threat of Stylometry
Stylometry is a potent threat to privacy. On raw, unaltered data, a Support Vector Machine (SVM) using bi-gram features can correctly identify a true user 82% of the time. Attackers utilize a wide array of features, from simple n-grams to character-level distributions and syntactic patterns, to build these linguistic fingerprints.
Methodology: Neural Generative Models
The authors propose using Neural Generative Models to create entirely synthetic corpora. Instead of releasing original tweets, they release "fake" tweets generated by a model trained on the originals.
The Multi-User Conditional Architecture
The core innovation is a structural modification to standard Recurrent Neural Networks (RNNs). The architecture is conditioned on a user-specific embedding ($e_j$). This allows the model to share a global understanding of English grammar and vocabulary while maintaining parameters that capture how a specific user uniquely constructs sentences.
The "Privacy Knob": Temperature ($\tau$)
The authors use a "temperature" parameter in the output layer (Softmax) to tune the model:
- Low Temperature: The model is "confident" and picks the most likely next character, producing text very similar to the original (High Utility, High Risk).
- High Temperature: The model becomes more "random," introducing noise that masks stylistic quirks (Low Risk, but potentially lower Utility).
Key Results and Insights
- Neural models outperform traditional obfuscation: While redaction and translation help, they don't offer a "tuning knob." Neural models allow data providers to choose a specific point on the risk-utility curve.
- Delta-RNN is the winner: The Delta-RNN variant generally offered the best trade-off, maintaining higher utility (especially in preserving word distributions) for a given level of privacy compared to the GRU or Elman RNN.
- Risk Reduction: All neural models significantly reduced re-identification risk. For example, at a temperature of 1.0, the Delta-RNN reduced the top-1 identification risk from 82% to roughly 44-51%.
- Utility Preservation: The models were remarkably good at preserving user-level sentiment scores and performed well in classification tasks (e.g., identifying tweets about specific topics), significantly outperforming redacted datasets.
Conclusion and Future Work
This work demonstrates that character-level neural synthesis is a viable path for privacy-preserving data release. By modeling at the character level, the models capture emoticons and unique punctuation vital for sentiment analysis while providing a mechanism to obfuscate identity. Future research will explore integrating formal $\epsilon$-differential privacy guarantees and scaling these models to even larger user populations.
Summary updated