SaLoRA: Safety-Alignment Preserved Low-Rank Adaptation
This addresses safety risks for model owners in personalized LLM fine-tuning, representing an incremental improvement over existing LoRA methods.
The paper tackles the problem that LoRA fine-tuning can compromise safety alignment in large language models, proposing SaLoRA to preserve safety while enabling targeted modifications, with experiments showing it outperforms other adapters-based approaches across various metrics.
As advancements in large language models (LLMs) continue and the demand for personalized models increases, parameter-efficient fine-tuning (PEFT) methods (e.g., LoRA) will become essential due to their efficiency in reducing computation costs. However, recent studies have raised alarming concerns that LoRA fine-tuning could potentially compromise the safety alignment in LLMs, posing significant risks for the model owner. In this paper, we first investigate the underlying mechanism by analyzing the changes in safety alignment related features before and after fine-tuning. Then, we propose a fixed safety module calculated by safety data and a task-specific initialization for trainable parameters in low-rank adaptations, termed Safety-alignment preserved Low-Rank Adaptation (SaLoRA). Unlike previous LoRA methods and their variants, SaLoRA enables targeted modifications to LLMs without disrupting their original alignments. Our experiments show that SaLoRA outperforms various adapters-based approaches across various evaluation metrics in different fine-tuning tasks.