Exploring Fine-Tuning Techniques for Removing
Tamper-Resistant Safeguards for Open-Weight LLMs
Name
zhang-sjzhang-meng-eecs-2025-thesis.pdf
Description
Thesis PDF
Size
2.67 MB
Format
Adobe PDF
Checksum (MD5)
c5cb0b16e30aa5ec98cf31a80515a96d
Author(s)
Zhang, Sarah
Advisor(s)
Kim, Yoon
Date Issued
February 2025
Publisher
Massachusetts Institute of Technology
Abstract
Open-source models present significant opportunities and risks, especially in dual-use scenarios where they can be repurposed for malicious tasks via adversarial fine-tuning. In this paper, we evaluate the effectiveness of Tampering Attack Resistance (TAR), a safeguard designed to protect against such adversarial attacks, by exploring its resilience to full-parameter and parameter-efficient fine-tuning. Our experiments reveal that while TAR enhances tamper resistance compared to models without safeguards, it remains susceptible to variability. Specifically, we observe inconsistencies where the same adversarial attack can succeed under some initializations and fail under others. This is a critical security risk as even a single instance of failure can lead to models being exploited for harmful purposes. These findings highlight the limitations of current tamper-resistant safeguards and emphasize the need for more robust safeguards to ensure the safe and ethical deployment of open-source models.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
