Langevin Unlearning: A New Perspective of Noisy Gradient Descent for Machine Unlearning

2024-01-18

arXiv.org

AI Summary

Addressing the challenge of implementing the “right to be forgotten” in machine learning—particularly for non-convex models, sequential or batch unlearning requests, and rigorous privacy guarantees—this paper proposes the Langevin Unlearning Framework. Our method unifies differentially private training with certified unlearning by integrating noisy stochastic gradient descent and Langevin dynamics, enabling provably private approximate certified unlearning. It supports non-convex optimization, balancing unlearning accuracy and model utility. Computationally, it achieves up to数十-fold speedup over full retraining for batch unlearning. Empirical evaluation on standard benchmarks demonstrates a three-way trade-off: effective unlearning (verified via membership inference and influence analysis), formal ε-differential privacy, and competitive generalization performance.

Technology Category

Machine Learning: PrivacyNatural Language Processing: Learning & Optimization for NLPSearch and Optimization: Learning to Search

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSecurity and Privacy: Security and privacy of machine learning and AI applicationsUser Modeling, Personalization and Recommendation: User privacy protection in personalized systems

Abstract

Machine unlearning has raised significant interest with the adoption of laws ensuring the ``right to be forgotten''. Researchers have provided a probabilistic notion of approximate unlearning under a similar definition of Differential Privacy (DP), where privacy is defined as statistical indistinguishability to retraining from scratch. We propose Langevin unlearning, an unlearning framework based on noisy gradient descent with privacy guarantees for approximate unlearning problems. Langevin unlearning unifies the DP learning process and the privacy-certified unlearning process with many algorithmic benefits. These include approximate certified unlearning for non-convex problems, complexity saving compared to retraining, sequential and batch unlearning for multiple unlearning requests.

Problem

• Machine Learning
• Right to be Forgotten
• Privacy and Efficiency

Innovation

• Langevin Unlearning
• Differential Privacy
• Noisy Gradient Descent