Collaborative Optimization of Label Protection in Large Language Models with an Adaptive Rényi Divergence Approach

Advancements in large language models (LLMs) are driving progress towards general artificial intelligence. Ensuring differential privacy during fine-tuning and inference on domain-specific data is essential, but balancing privacy and utility, especially for sequential and label data, remains challenging. Existing methods protect sequential data well but struggle with label data due to traditional definitions of label differential privacy, which ignore label-feature correlations and weaken label privacy. We propose an optimized privacy-utility trade-off for label data by introducing a new definition for conditional feature-label differential privacy and enhancing the double randomized response (DRR) mechanism with a Rényi divergence optimizer. Additionally, a dynamic perturbation factor function improves usability across various tasks. Experimental results show that our method enhances resistance to label inference attacks by over 20% while maintaining similar accuracy, offering stronger privacy with a reduced privacy overhead and achieving an optimal balance during LLMs fine-tuning and inference.