Researchers have developed a new method called BadMoE to exploit vulnerabilities in Mixture-of-Experts (MoE) large language models. This technique targets the routing mechanism of MoE architectures, which selectively activates subsets of experts for processing inputs. BadMoE identifies underutilized experts and uses specially crafted triggers to manipulate their activation, thereby injecting predefined backdoor behaviors into the model. The attack has been demonstrated on models like Mixtral, OLMoE, and Deepseek across various tasks, including classification and targeted refusal. AI
IMPACT Highlights a new class of vulnerabilities in MoE models, potentially impacting their security and trustworthiness in deployment.
RANK_REASON Academic paper detailing a new attack method on LLM architectures. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
