Reproduce “GateBreaker: Gate-Guided Attacks on Mixture-of-Expert LLMs” 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
Lichao Wu, Sasha Behrouzi, Mohamadreza Rostami, Stjepan Picek, Ahmad-Reza Sadeghi
The GateBreaker framework introduces a white-box attack on Mixture-of-Experts (MoE) Large Language Models (LLMs) to compromise their safety alignment. This method achieved an average attack success rate of 64.9% across eight state-of-the-art MoE LLMs and 60.9% on MoE Vision Language Models, by disabling a mere 2.6% of neurons per targeted expert layer with negligible impact on model utility and demonstrating strong cross-model transferability.
Key insights
- Safety alignment in MoE LLMs is not uniformly distributed but highly concentrated within a small, identifiable subset of neurons across specific experts and layers, particularly in the gate-projection layers.
- Fine-grained, neuron-level intervention can effectively compromise safety alignment without significantly degrading the model's general utility, indicating that safety mechanisms can be decoupled from core capabilities.
- Structural vulnerabilities in safety alignment are transferable across different models within the same architectural family (e.g., sibling variants), suggesting common design patterns for safety mechanisms that can be exploited.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.