MoE LLM vulnerability exploited by 'RepetitionCurse' DoS attack

Researchers have identified a vulnerability in Mixture-of-Experts (MoE) Large Language Models that can be exploited as a denial-of-service attack. Adversarial inputs can cause the model's router to concentrate all processing on a small subset of experts, creating bottlenecks and increasing inference latency. The proposed 'RepetitionCurse' method uses simple repetitive token patterns to trigger this imbalance, significantly degrading model performance and availability. AI

IMPACT This research highlights a critical security vulnerability in MoE architectures, potentially impacting the reliability and availability of deployed LLM services.

RANK_REASON The cluster contains a research paper detailing a new attack method against MoE LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• LLMs
• Mixtral-8x7B
• Mixture-of-Experts
• RepetitionCurse
• Ruixuan Huang

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →