RESEARCHInvestigateNEXT 12 MONTHS
Memorization Is Not Extraction: Tight Differential-Privacy Bounds and Audit Blind Spots
arXiv cs.CL — Computation and Language
Factual evidence
What the source reports
Research proves mathematical differential privacy bounds do not prevent adaptive extraction of memorized training data in LLMs.