Scholarlink

To effectively deploy Large Language Models (LLMs) in application-specific settings, fine-tuning techniques are applied to enhance performance on specialized tasks. This process often involves fine-tuning on user data, which may contain sensitive information. Although not recommended, it is not uncommon for users to send passwords in messages, and fine-tuning models on this could result in passwords being leaked. In this study, a Large Language Model is fine-tuned with customer support data and passwords from the RockYou password wordlist using Low-Rank Adaptation (LoRA). RockYou is selected as it is one of the most well-known passwords and is including in most Kali Linux distributions. Out of the first 200 passwords from the list, 37 were successfully recovered. Further, causal tracing is used to identify that password information is largely located in a few layers. Lastly, Rank One Model Editing (ROME) is used to remove the password information from the model, resulting in the number of passwords recovered going from 37 to 0.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leaking LoRa: An Evaluation of Password Leaks and Knowledge Storage in Large Language Models

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年

查看全文
  • Ryan Marinelli,
  • Magnus Eckhoff

摘要

To effectively deploy Large Language Models (LLMs) in application-specific settings, fine-tuning techniques are applied to enhance performance on specialized tasks. This process often involves fine-tuning on user data, which may contain sensitive information. Although not recommended, it is not uncommon for users to send passwords in messages, and fine-tuning models on this could result in passwords being leaked. In this study, a Large Language Model is fine-tuned with customer support data and passwords from the RockYou password wordlist using Low-Rank Adaptation (LoRA). RockYou is selected as it is one of the most well-known passwords and is including in most Kali Linux distributions. Out of the first 200 passwords from the list, 37 were successfully recovered. Further, causal tracing is used to identify that password information is largely located in a few layers. Lastly, Rank One Model Editing (ROME) is used to remove the password information from the model, resulting in the number of passwords recovered going from 37 to 0.