New attack exploits MoE router telemetry for privacy leakage

Researchers have developed a new membership inference attack that exploits telemetry data from Mixture-of-Experts (MoE) language models. This attack, called a router-augmented membership inference attack, combines traditional output signals with aggregated routing features to determine if a specific data example was used during the model's fine-tuning process. The findings indicate that this telemetry data consistently improves the accuracy of membership inference attacks across various MoE architectures and data domains, even when privacy-preserving measures are implemented. AI

IMPACT Highlights a new privacy vulnerability in MoE models, potentially impacting how telemetry data is handled and secured.

RANK_REASON Academic paper detailing a new privacy attack method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Hugging Face
• LoRA
• Membership inference attack
• Mixture-of-Experts

• paper
• safety

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