Authors - Hiep Nghia Phan, Duy Nguyen Ngoc Abstract - Low-Rank Adaptation (LoRA) is widely used for parameter-efficient fine-tuning of large language models because it is lightweight, modular, and easy to distribute. However, the growing practice of sharing third-party LoRA modules also creates security concerns. A malicious adapter can introduce hidden behaviors, backdoors, or other risks while appearing to function normally. Existing research has largely focused on detecting whether a LoRA module is malicious. In practice, deployment decisions often require a more nuanced assessment of risk. This paper presents a measurable framework that evaluates LoRA security across four dimensions: supply-chain integrity, static weight characteristics, dynamic behavior, and deployment-time observations. The resulting indicators are normalized and combined into a composite risk score that supports comparison and prioritization of LoRA modules. The framework was evaluated using benign and backdoored LoRA modules attached to a frozen base language model. The results show a clear separation between the two groups even when their task performance remains similar. Dynamic behavioral testing and static weight analysis contribute the most useful signals, while deployment-time monitoring provides additional evidence of long-term operational risk. The proposed framework provides a practical mechanism for integrating security assessment into LoRA selection, governance, and deployment workflows.
Open Zoom