Xun Wang

Xun Wang

Ph.D student at CISPA

Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks through prompt engineering and fine-tuning. However, transferring knowledge between different LLMs while preserving privacy remains a significant challenge. This work introduces POST (Privacy-preserving Soft prompt Transfer), a novel framework that enables efficient transfer of soft prompts between LLMs while maintaining strong privacy guarantees.

Key Contributions

  • Privacy-Preserving Transfer: First framework to enable soft prompt transfer with formal differential privacy guarantees
  • Efficiency Optimization: Significantly reduces computational overhead compared to traditional fine-tuning approaches
  • Cross-Model Compatibility: Enables knowledge transfer between different LLM architectures and sizes
  • Theoretical Analysis: Provides privacy and utility guarantees with rigorous theoretical foundations

Methodology

POST combines differential privacy mechanisms with knowledge distillation techniques to:

  1. Extract transferable knowledge from source prompts while adding calibrated noise
  2. Compress soft prompt representations into privacy-preserving embeddings
  3. Adapt transferred knowledge to target LLMs through efficient fine-tuning

Results

  • Privacy: Achieves ε-differential privacy with ε < 1.0 across all experiments
  • Efficiency: 10x faster than full model fine-tuning with comparable performance
  • Utility: Maintains 95%+ of original performance on downstream tasks
  • Generalization: Successfully transfers across different model families (GPT, LLaMA, T5)

Impact

This work addresses critical challenges in LLM deployment where privacy and efficiency are paramount, opening new possibilities for federated learning and collaborative AI development without compromising sensitive information.

Recommended citation: Wang, X., et al. (2025). "Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs." International Conference on Machine Learning (ICML).
Download Paper