Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

GenBFA: An Evolutionary Optimization Approach to Bit-Flip Attacks on LLMs

Created by
  • Haebom

Author

Sanjay Das, Swastik Bhattacharya, Souvik Kundu, Shamik Kundu, Anand Menon, Arnab Raha, Kanad Basu

Outline

In this paper, we study the vulnerability of large-scale language models (LLMs) to hardware-based threats, especially bit-flip attacks (BFA). While previous works claim that transformer-based architectures are more robust against BFA, we show that even a few bit-flips can severely degrade the performance of LLMs with billions of parameters. To address this issue, we propose AttentionBreaker, a novel framework that efficiently explores the parameter space of LLMs to identify important parameters, and GenBFA, an evolutionary optimization strategy that finds the most significant bits. Experimental results demonstrate that AttentionBreaker exposes a serious vulnerability of LLMs, where even a few bit-flips can completely collapse the model performance.

Takeaways, Limitations

•
Takeaways:
◦
Empirically shows that LLM is very vulnerable to BFA.
◦
We propose AttentionBreaker, a novel framework for efficiently exploring the parameter space of LLM.
◦
We propose GenBFA, an evolutionary optimization strategy to efficiently find important bits.
◦
We highlight that even a small number of bit flips can dramatically degrade LLM performance.
◦
Highlighting the importance of LLM security enhancement research.
•
Limitations:
◦
The effectiveness and efficiency of AttentionBreaker and GenBFA may vary depending on the specific LLM and BFA attack methods.
◦
Further research is needed on various LLM architectures, sizes, and quantization methods.
◦
Further analysis of the attack success rate and impact in real-world environments is needed.
◦
There is a lack of research on defense techniques against proposed attack techniques.
👍
Made with Slashpage