ICML 2026 Paper Proposes ALU Framework for Large-Scale Machine Unlearning

Researchers presented the paper "Unlearning with Asymmetric Sources" at ICML 2026 in Seoul, introducing the Asymmetric Langevin Unlearning (ALU) framework. The study aims to address the tradeoff between privacy certification and model utility in machine unlearning, enabling models to "forget" specific training data without compromising overall performance.

Background and Limitations of Existing Methods

Machine unlearning seeks to allow models to forget specific training data. However, existing noise-based certified machine unlearning methods have a fundamental limitation: the noise added to guarantee unlearning certification severely degrades model utility in large-scale deletion scenarios, creating an insurmountable "utility ceiling." Furthermore, previous studies often assumed perfect alignment between public and private data. In contrast, this team explicitly modeled and quantified the impact of real-world "distribution mismatch" on unlearning efficacy, making the theory more applicable to practical deployment.

Core Advantages of the ALU Framework

To overcome these bottlenecks, the researchers proposed the ALU framework. Its core idea is to utilize public data to reduce the privacy cost of machine unlearning. Mathematical proofs show that injecting public data can improve the suppression factor of the unlearning cost to O(1/n_pub^2). Thanks to this mechanism, ALU not only supports large-scale unlearning of a constant proportion of data but also theoretically outperforms retraining models from scratch in computational complexity, while still maintaining strict unlearning certification.

2026-07-07 ~ 2026-07-07 · 5 related posts