Researchers have developed a new set of attacks called LeakyLMs that can infer proprietary information about language models, including their architecture and inference optimizations, by analyzing token generation timing. This method can even extract details through remote APIs. One attack successfully identified that Google Gemini Flash 2.5 uses speculative decoding with a draft context window of approximately 128K tokens. Another attack can recover architectural properties like the number of transformer layers, hidden dimension size, and attention heads, achieving over 90% accuracy in identifying Llama model configurations. AI
IMPACT Reveals potential security vulnerabilities in deployed LLMs, impacting how models are secured and how their architectures are protected.
RANK_REASON Academic paper detailing a new attack method on language models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
