New unlearning method uses per-instance bounds to reduce noise

Researchers have developed a new method for certified machine unlearning that uses per-instance sensitivity bounds to calibrate noise injection. This approach aims to reduce the performance degradation often seen with traditional methods that use worst-case sensitivity. The study derives high-probability per-instance sensitivity bounds for ridge regression trained via Langevin dynamics, demonstrating certified unlearning with significantly less noise. Experiments in linear settings and empirical evidence in deep learning settings support the theoretical findings. AI

IMPACT This research offers a more efficient method for unlearning data from machine learning models, potentially improving privacy and reducing performance loss.

RANK_REASON This is a research paper detailing a new theoretical approach and experimental validation for machine unlearning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• arXiv
• Hanna Benarroch

• paper
• safety

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