TL;DR
This paper introduces WASP, a novel framework for generating high-quality, differentially private synthetic data, especially useful when real private data is scarce. WASP tackles issues of noisy synthetic data and pre-trained model bias by using a weighted fusion of multiple language models and a contrastive learning approach.
What Problem Does This Paper Address?
In the age of AI, high-quality training data is crucial, but often sensitive private data is required. Generating synthetic data that mimics real data while guaranteeing Differential Privacy (DP) is a promising solution. However, existing methods struggle with:
- Data Scarcity: They require a large number of private samples to work effectively, which isn't always available.
- Noisy Synthetic Data: The generated data can contain errors or be of low quality, hindering downstream model performance.
- Pre-trained Model Bias: Relying on a single pre-trained model can introduce its own biases and may not be optimal for all tasks.
WASP aims to overcome these limitations by enabling collaborative synthesis using multiple pre-trained language models (PLMs) even with limited private data.
Methods
WASP employs a multi-stage, iterative approach:
- Weighted Parallel Data Synthesis: Multiple PLMs generate synthetic data in parallel. Each PLM's contribution is dynamically weighted based on its performance in previous rounds, addressing model bias.
- Differentially Private Top-Q Voting: To improve the estimation of the real data distribution from limited private samples, WASP uses a "Top-Q" voting mechanism. Each private sample votes for the Q nearest and Q furthest synthetic samples, providing a more robust signal than simple voting. This process is protected by DP.
- Sample Selection & Contrastive In-context Learning (ICL): Based on the voting results, WASP identifies both high-quality (nearest to private data) and low-quality (furthest from private data) synthetic samples. These are then used to construct a "contrastive prompt" for the next generation round. This prompt guides PLMs to generate data that is more similar to high-quality samples and less similar to low-quality ones, effectively reducing noise.
- PLM Importance Weighting: The voting results are used to calculate weights for each PLM, prioritizing those that generate samples closer to the real data distribution. This ensures that more capable models contribute more to the final synthetic dataset.
The framework is designed to be iterative, refining the synthetic dataset over multiple rounds.
Datasets
The paper evaluates WASP on six well-defined natural language processing tasks:
- IMDb (movie review sentiment analysis)
- Yelp-Category (business review field classification)
- Yelp-Rating (business review rating classification)
- Openreview-Category (paper review research area classification)
- Openreview-Rating (paper review rating classification)
- Banking (online banking queries field classification)
These datasets are standard benchmarks for NLP tasks. The authors use subsets of these datasets for training the synthetic data generation process and for fine-tuning downstream Small Task-specific Models (STMs).
Key Insights and Ideas
- Collaboration is Key: Fusing multiple PLMs, even those with varying strengths, leads to better synthetic data than relying on a single model, especially in data-scarce scenarios.
- Contrastive Learning for Quality: Using both high-quality and low-quality synthetic samples as feedback during generation is crucial for reducing noise and improving the overall quality of the synthetic dataset. It teaches the models what to emulate and what to avoid.
- Dynamic Weighting Mitigates Bias: Dynamically adjusting the contribution of each PLM based on its performance in previous rounds effectively addresses the bias inherent in individual models.
- Top-Q Voting for Robustness: Extending simple voting to a Top-Q mechanism with decaying weights allows for a more accurate estimation of the private data distribution from limited samples.
Key Results
- Superior Performance: WASP consistently outperforms existing baselines (like Aug-PE and Pre-Text) across various downstream tasks and datasets, demonstrating significant improvements in STM accuracy. For example, on the IMDb dataset, WASP achieved an accuracy of 89.52%, compared to Aug-PE's best of 89.00%.
- Effectiveness with Limited Data: WASP shows remarkable performance even with as few as 100 private samples, a scenario where other methods struggle.
- PLM Agnostic: The framework's performance is not heavily dependent on selecting specific "best" PLMs, making it more practical.
- Scalability: WASP scales well with the number of PLMs (K) and maintains high performance even under tighter privacy budgets (lower epsilon values).
- Federated Learning Compatibility: WASP can be effectively extended to federated learning settings, maintaining its superiority over baselines like Pre-Text.
Limitations and Future Work
- Computational Cost: While WASP is efficient in terms of API calls, the iterative nature and use of multiple PLMs can still be computationally intensive.
- Sample-Level DP Focus: The primary focus is on sample-level DP. While user-level DP is discussed as an extension, further exploration might be beneficial.
- Future Directions: The authors suggest exploring more precise sample-level weighting or selection mechanisms to further enhance DP synthetic data quality. They also propose verifying WASP's effectiveness on non-classification tasks.
Summary updated