New formula calibrates DP-SGD noise against membership inference attacks

Researchers have developed a new closed-form formula to calibrate noise levels for DP-SGD, a method used to protect training data by adding noise to gradients. This formula specifically applies to DP-SGD with random allocation, where each record is used once per epoch at a random step. The derived formula provides an upper bound on the accuracy of membership inference attacks (MIAs) against models trained with this method, offering a more efficient way to determine the necessary noise multiplier compared to traditional numerical privacy accountants. AI

IMPACT Provides a more efficient method for calibrating noise in DP-SGD, potentially leading to better privacy-utility trade-offs in machine learning models.

RANK_REASON The cluster contains a research paper detailing a new mathematical formula for privacy-preserving machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• DP SGD
• Gaussian mixture model
• Gaussian noise
• gradients
• Hugging Face
• membership inference attack
• Murat Bilgehan Ertan
• Gaussian distribution
• chi-squared divergence

• paper
• safety

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