New N-Gram attack probes black-box LLMs for training data leakage

Researchers have developed a new membership inference attack called N-Gram Coverage Attack, which can be used on black-box language models like GPT-4 by only analyzing their text outputs. This method leverages the observation that models tend to memorize and regenerate text patterns from their training data. The attack demonstrates strong performance, even rivaling white-box methods, and its effectiveness increases with more generated sequences. Notably, the study found that newer models like GPT-4o show improved resistance to such attacks, indicating enhanced privacy measures. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT New black-box attack method could challenge privacy protections in API-only models, though newer models show improved robustness.

RANK_REASON The cluster contains an academic paper detailing a new method for membership inference attacks on language models.

Read on arXiv cs.CL →

• N-Gram Coverage Attack
• GPT-4
• OpenAI
• GPT-4o
• Skyler Hallinan

• paper
• safety