New 'structural memorization' vulnerability found in MoE LLMs

Researchers have identified a new vulnerability in Mixture-of-Experts (MoE) large language models called "structural memorization." This vulnerability allows models to memorize not only their training data but also the specific internal pathways used to process that data, creating a unique routing fingerprint. This characteristic can be exploited by adversaries to recover memorized expert pathways, even without access to the original training data, by analyzing unconditional generation outputs. The findings are particularly concerning for distributed deployments like federated learning and edge-cloud systems, where routing information is exposed across multiple trust boundaries. AI

IMPACT This research highlights potential security risks in distributed LLM deployments, necessitating new detection methods for compromised models.

RANK_REASON The cluster contains an academic paper detailing a new vulnerability in LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• backdoor injection
• edge-cloud systems
• federated learning
• Membership inference attack
• mixture of experts
• router manipulation
• structural memorization

• paper
• safety

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