New method improves privacy auditing for ML models despite data distribution shifts

Researchers have developed a new method called "shift-aware score transport" to improve the accuracy of privacy auditing for machine learning models. This technique addresses the challenge of distribution shift, where the data used for auditing differs from the model's training data, leading to skewed privacy assessments. By learning the relationship between data representations and privacy scores from out-of-distribution non-members, the method can transport these insights to the target distribution and correct for audit errors. This approach yields more stable privacy estimates and tighter lower bounds compared to existing methods, particularly in observational auditing settings where model retraining is not feasible. AI

IMPACT Enhances the reliability of privacy assessments for deployed AI models, particularly when retraining is not an option.

RANK_REASON The cluster contains a research paper detailing a new methodology for privacy auditing in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Membership Inference Attacks
• out-of-distribution (OOD) non-members
• privacy auditing
• ScienceCast
• Shift-aware score transport

• paper
• safety

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