TL;DR
This paper introduces DP-ST, a novel method for generating private text documents under local differential privacy (LDP) guarantees. By decomposing documents into semantic triples and using a neighborhood-aware privatization approach, DP-ST achieves a better balance between privacy and utility, even at lower privacy budgets, compared to existing methods.
What Problem Does This Paper Address?
The core challenge in text privatization is the "privacy-utility tradeoff." Existing methods for achieving differential privacy (DP) in natural language processing (NLP) often struggle to maintain text coherence and utility, especially under strict local DP (LDP) guarantees. This is particularly true for document-level privacy, where making an entire document indistinguishable from others requires very high privacy budgets (epsilon, ɛ), rendering the privatized text often nonsensical or grammatically incorrect. Previous approaches have either focused on word-level privacy, leading to incoherent documents, or document-level privacy with less control over what information is protected, demanding high ɛ values. This paper aims to bridge this gap by offering a method that provides strong privacy guarantees while preserving text quality and coherence, even with more relaxed privacy budgets.
Methods
DP-ST employs a "divide-and-conquer" strategy centered around semantic triples (subject-verb-object relationships). The process involves three main stages:
- Decomposition into Semantic Triples: Input documents are first broken down into SVO triples using the Stanford Open Information Extraction (OpenIE) tool. Post-processing steps, including Locality-Sensitive Hashing (LSH) and perplexity filtering (using a GPT-2 model), are applied to refine these triples and select the most coherent ones.
- Neighborhood-Aware DP Privatization: A public corpus of semantic triples is built from a large text corpus (FineWeb). This corpus is then clustered using k-means. When privatizing an input triple, DP-ST considers only the public triples within the most semantically similar cluster (the "privatization neighborhood"). The Exponential Mechanism (EM) is then used to select a privatized triple from this neighborhood, with a privacy budget (ɛ) distributed across all triples in a document. This approach is termed "neighborhood-aware DP," relaxing the strict LDP by limiting the scope of privacy guarantees to within a cluster.
- Document Reconstruction with LLMs: The privatized triples are then fed into a Large Language Model (LLM) to reconstruct a coherent private document. This leverages the "post-processing principle" of DP, ensuring that the LLM's generation process does not compromise the privacy guarantees already established.
Datasets
The authors evaluated DP-ST on five diverse datasets:
- Reuters News: A 50-class authorship identification task.
- Spooky Authors: An adversarial authorship identification task.
- Reddit Mental Health: A dataset focused on mental health discussions.
- Trustpilot Reviews: Used for both adversarial gender identification and sentiment analysis.
- Yelp Reviews: Used for authorship identification and sentiment analysis.
No new benchmarks were introduced; instead, existing datasets with privacy-sensitive tasks were utilized.
Key Insights and Ideas
- Semantic Triples as the Core Unit: Focusing privatization on semantic triples allows for a more granular approach to privacy, targeting the essential information while reducing the overall privacy budget needed.
- Neighborhood-Aware DP is Key: Restricting the privatization scope to semantically similar triples within a cluster significantly improves the privacy-utility tradeoff compared to global LDP. This is because the search space for privatization is drastically reduced.
- LLM Reconstruction is Crucial for Coherence: The LLM post-processing step is vital for transforming privatized triples back into natural and coherent text, a significant improvement over methods that directly output privatized words or sentences.
- Smaller Clusters = Better Tradeoffs: The results suggest that a smaller number of clusters (leading to larger privatization neighborhoods) generally yields better relative gains, indicating a more favorable privacy-utility balance.
- Model Size Paradox: While larger LLMs for reconstruction might improve semantic similarity, they can sometimes weaken empirical privacy gains, suggesting a complex interplay between model choice and privacy.
Key Results
- DP-ST consistently outperformed state-of-the-art methods in achieving a positive privacy-utility tradeoff across multiple datasets.
- It demonstrated strong coherence and semantic similarity preservation, even at lower ɛ values.
- DP-ST achieved positive relative gains (indicating privacy benefits outweighing utility loss) on average, a feat not matched by most comparative methods.
- The method's computational efficiency was comparable to existing methods, especially when using smaller LLMs for reconstruction.
Limitations and Future Work
- Reliance on Public Corpus: The quality and expressiveness of the privatized triples depend on the completeness of the public triple corpus. The current corpus, while large, might limit semantic expressiveness.
- OpenIE Limitations: The use of OpenIE for triple extraction, while efficient, is not state-of-the-art and has inherent limitations in capturing all semantic nuances. Future work could explore more advanced extraction techniques.
- Fallback Mechanism: If no triples are extracted from a text, DP-ST currently returns the original text, which could be a privacy concern for shorter texts.
- Comparability of DP Notions: The paper acknowledges that directly comparing methods with different DP notions (MDP, ɛ-DP, (ɛ, δ)-DP) is challenging, even when using uniform document-level budgets.
- LLM Data Contamination: While minimized by not feeding original texts to LLMs, the possibility of LLMs having seen evaluation data during pre-training is a potential concern.
- Future Work: The authors suggest further research into the feasibility of the divide-and-conquer paradigm for DP text privatization and exploring more advanced post-processing techniques to enhance privacy and utility. They also highlight the need for a more rigorous analysis of the impact of LLM parameters and cluster size on DP-ST's performance.
Summary updated