Reproduce “TFL: Targeted Bit-Flip Attack on Large Language Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingkai Guo, Chaitali Chakrabarti, Deliang Fan
Researchers at Arizona State University introduced TFL, a targeted bit-flip attack framework for large language models, demonstrating the ability to manipulate LLM outputs to generate specific false answers for chosen prompts while maintaining high performance on unrelated tasks. The method achieves this with as few as 4 bit flips, significantly outperforming existing untargeted bit-flip attacks.
Key insights
- LLMs are vulnerable to highly targeted, stealthy manipulation with an exceptionally low number of bit flips, challenging assumptions of robustness based on model size and redundancy.
- The success and stealth of targeted bit-flip attacks are significantly enhanced when using keywords that are semantically or factually relevant to the desired malicious output.
- Balancing the targeted attack objective with the preservation of benign model behavior through metrics like the Aux Utility Score is crucial for achieving stealth and preventing catastrophic model degradation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.