Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

Defending MoE LLMs against Harmful Fine-Tuning via Safety Routing Alignment

Created by
  • Haebom

Author

Jaehan Kim, Minkyoo Song, Seungwon Shin, Sooel Son

Outline

This paper analyzes vulnerabilities in the security of large-scale language models (LLMs) using the Mixture-of-Experts (MoE) architecture and proposes SafeMoE, a novel secure fine-tuning method to address these vulnerabilities. Specifically, we highlight the problem that routing decisions for harmful inputs fluctuate significantly after fine-tuning, making them vulnerable to harmful fine-tuning (HFT) attacks. SafeMoE mitigates routing fluctuations by penalizing the difference between the routing weights of the initially safe-aligned model and the fine-tuned model, thereby maintaining security. Experimental results demonstrate that SafeMoE effectively mitigates HFT attacks and outperforms existing defense methods with minimal degradation in operational utility.

Takeaways, Limitations

•
Takeaways:
◦
We raised a critical issue regarding HFT attacks, a security vulnerability of MoE LLM, and proposed an effective defense technique called SafeMoE to address it.
◦
SafeMoE maintains security by directly mitigating routing fluctuations and outperforms existing defense methods.
◦
We demonstrate the effectiveness of SafeMoE across a range of model sizes through experiments on open-source MoE LLM.
•
Limitations:
◦
Although this paper demonstrates the effectiveness of SafeMoE on a specific open-source model, further research is needed to determine its generalizability to other architectures and model sizes.
◦
Additional studies may be required to determine the optimal hyperparameter settings for SafeMoE and to analyze its robustness against various HFT attack types.
◦
There is a computational overhead (2%), and considering this, research is needed on ways to minimize performance degradation when training large-scale models.
👍
Made with Slashpage