New metric-normalized posterior leakage (mPL) enhances privacy for joint AI consumption

Researchers have developed a new privacy metric called Metric-Normalized Posterior Leakage (mPL) to address limitations in existing differential privacy methods, particularly for machine learning systems used under joint observation. mPL measures the shift in posterior odds induced by data releases, offering a more accurate privacy guarantee when multiple data points are analyzed together. The proposed Adaptive mPL (AmPL) framework operationalizes this by perturbing data, using a learned attacker for auditing, and adapting parameters to balance privacy and utility, as demonstrated in a word-embedding case study. AI

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

IMPACT Introduces a more robust privacy metric for ML systems, potentially improving data protection in joint consumption scenarios.

RANK_REASON Academic paper introducing a new privacy metric and framework for machine learning systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Metric-Normalized Posterior Leakage
• mPL
• Metric differential privacy
• mDP
• local differential privacy
• LDP
• Adaptive mPL
• AmPL
• arXiv

• paper
• safety