Abstract
Ensuring strong theoretical privacy guarantees on text data is a challenging problem which is usually attained at the expense of utility. To improve the practicality of privacy-preserving text analyses, we propose PRIVEMB, a release mechanism that takes any embedding vector as input and releases a corresponding private vector satisfying $(\epsilon, \delta)$-Lipschitz privacy—a metric-space extension of Differential Privacy. By combining random dimensionality reduction with calibrated noise addition, PRIVEMB overcomes the "curse of dimensionality" that typically degrades the utility of high-dimensional private vectors. Empirical experiments on multiple word embedding models and NLP datasets demonstrate that PRIVEMB achieves significant utility gains, in some cases exceeding 10-20% over existing state-of-the-art techniques.
Introduction: The Challenge of High-Dimensional Privacy
Text embeddings (e.g., GloVe, FastText) are essential for modern NLP but are vulnerable to privacy attacks such as embedding inversion and membership inference. Standard Differential Privacy (DP) approaches often fail here because the required noise magnitude typically scales with the vector dimensionality ($d$). In high-dimensional spaces, this noise "drowns out" the signal, making private vectors ineffective for downstream tasks like sentiment analysis.
The PRIVEMB Mechanism
The PRIVEMB algorithm addresses this utility-privacy tradeoff through a two-stage process:
- Random Projection: The high-dimensional embedding is projected into a lower-dimensional space ($m$) using a Gaussian random matrix. This preserves the geometric structure of the data while reducing the sensitivity of the release mechanism.
- Calibrated Noise Addition: Noise is added in the projected space. Because the dimensionality is lower, the required noise magnitude is significantly reduced compared to adding noise in the original space.
PRIVEMB satisfies $(\epsilon, \delta)$-Lipschitz privacy, which acknowledges the underlying Euclidean metric space of embeddings. This ensures that similar words remain similarly private, preserving the relative distances crucial for NLP tasks. Furthermore, because privacy is applied at the point of release, the mechanism allows for infinite composition: researchers can train multiple models on the private vectors without further depleting the privacy budget.
Empirical Evaluation
We evaluated PRIVEMB (M2) against a baseline (M1) that adds noise in the original high-dimensional space. Tests were conducted using GloVe and FastText embeddings across various classification tasks (MR, CR, TREC-6, etc.) from the SentEval suite.
Key Results
PRIVEMB consistently demonstrated superior preservation of geometric properties and higher classification accuracy.
As shown in the distance approximation results, M2 significantly reduces the error in preserving distances between word pairs. We found that more aggressive dimensionality reduction (higher $\beta$) often improved utility, as the benefit of reduced noise outweighed the distortion from the projection.
Inner product preservation followed a similar trend, with M2 maintaining the geometric relationships necessary for high-performance NLP. These improvements led to substantial gains in downstream accuracy: on the TREC-6 question classification task, M2 achieved a 20% gain over M1 at $\epsilon=10$. On the CR (Customer Reviews) dataset, M2 achieved a 7% gain.
Conclusion
PRIVEMB provides a robust path forward for the private release of text embeddings. By leveraging random projections to mitigate the impact of high dimensionality, it offers a more favorable utility-privacy tradeoff than traditional methods. While a performance gap still exists relative to non-private baselines, the theoretical and empirical results presented here mark a significant advancement in practical privacy-preserving NLP.
Summary updated