New framework unifies privacy attacks, introduces Bayesian approach

Researchers have unified three leading membership inference attacks (MIAs) – LiRA, RMIA, and BASE – under a single exponential-family log-likelihood ratio framework. This unification reveals a hierarchy of model complexity, connecting RMIA and LiRA as endpoints. The study also introduces BaVarIA, a Bayesian variance inference attack designed to address performance bottlenecks at small shadow-model budgets. BaVarIA offers stable performance across various testbeds and budgets, outperforming LiRA, particularly in low-shadow-model and offline scenarios. AI

IMPACT This research provides a unified understanding of privacy attacks and introduces a more robust method for auditing model privacy, potentially improving security practices.

RANK_REASON The cluster contains an academic paper detailing a new theoretical framework and attack method for machine learning privacy. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• BaVarIA
• Carlini et al.
• Lassila et al.
• LiRA
• Rickard Brännvall
• RMIA
• Zarifzadeh et al.

• paper
• safety

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