This research introduces a novel approach to generate synthetic data that maintains the utility of real data while ensuring differential privacy protection.

• Leverages multiple pre-trained language models through weighted fusion techniques
• Employs contrastive learning to improve quality in data-deficient scenarios
• Provides formal privacy guarantees through differential privacy mechanisms
• Demonstrates practical applications for sensitive data environments

Key significance: Organizations can now create high-quality training datasets without compromising individual privacy, especially crucial for sensitive domains like healthcare and finance.

Contrastive Private Data Synthesis via Weighted Multi-PLM Fusion