Differential privacy (DP) has been proven to be an effective universal solution for privacy protection in language models. Nevertheless, the introduction of DP incurs significant computational overhead. One promising approach to this challenge is to integrate Parameter Efficient Fine-Tuning (PEFT) with DP, leveraging the memory-efficient characteristics of PEFT to reduce the substantial memory consumption of DP. Given that fine-tuning aims to quickly adapt pretrained models to downstream tasks, it is crucial to balance privacy protection with model utility to avoid excessive performance compromise. In this paper, we propose a Relevance-based Adaptive Private Fine-Tuning (Rap-FT) framework, the first approach designed to mitigate model utility loss caused by DP perturbations in the PEFT context, and to achieve a balance between differential privacy and model utility. Specifically, we introduce an enhanced layer-wise relevance propagation process to analyze the relevance of trainable parameters, which can be adapted to the three major categories of PEFT methods. Based on the relevance map generated, we partition the parameter space dimensionally, and develop an adaptive gradient perturbation strategy that adjusts the noise addition to mitigate the adverse impacts of perturbations. Extensive experimental evaluations are conducted to demonstrate that our Rap-FT framework can improve the utility of the fine-tuned model compared to the baseline differentially private fine-tuning methods, while maintaining a comparable level of privacy protection.

Large Language Models (LLMs) have shown incomparable representation and generalization capabilities, which have led to significant advancements in Natural Language Processing (NLP). Before deployment, the pre-trained LLMs often need to be tailored to specific downstream tasks for improved performance, which is commonly referred to as downstream alignment. This is a costly effort considering the needed manpower, training resources, and downstream-specific data. While much attention has been paid to protecting the copyright of the models themselves, the copyright protection of LLM alignment has been largely overlooked. In this paper, we present Watermark Embedding for Downstream Alignment (WEDA) scheme, which can provide effective copyright protection for two popular LLM alignment techniques parameter-efficient fine-tuning (PEFT) and in-context learning (ICL). For alignment through PEFT, we propose a Chain of Thought (CoT) based solution to embed watermarks into the PEFT weights. Furthermore, we extend this solution to safeguard alignment through ICL by utilizing the prefix-integrated CoT to watermark examples embedded within ICL prompts. We conduct an extensive experimental evaluation to demonstrate the effectiveness of our proposed scheme.