Machine Unlearning under Retain-Forget Entanglement

2026-03-27

AI Summary

This work addresses the challenge in machine unlearning where forgetting samples are often semantically or feature-wise entangled with retained samples, leading to degraded performance on the latter. To tackle this “retain–forget entanglement,” the paper proposes the first two-stage optimization framework: in the first stage, an augmented Lagrangian method is employed to maximize loss on the forget set while safeguarding unrelated retained samples; in the second stage, a Wasserstein-2 distance–regularized gradient projection is introduced to mitigate performance degradation on semantically related retained samples. Extensive experiments across multiple benchmark datasets and model architectures demonstrate that the proposed method achieves superior balance between effective forgetting and retention performance, significantly outperforming existing approaches.

Abstract

Forgetting a subset in machine unlearning is rarely an isolated task. Often, retained samples that are closely related to the forget set can be unintentionally affected, particularly when they share correlated features from pretraining or exhibit strong semantic similarities. To address this challenge, we propose a novel two-phase optimization framework specifically designed to handle such retai-forget entanglements. In the first phase, an augmented Lagrangian method increases the loss on the forget set while preserving accuracy on less-related retained samples. The second phase applies a gradient projection step, regularized by the Wasserstein-2 distance, to mitigate performance degradation on semantically related retained samples without compromising the unlearning objective. We validate our approach through comprehensive experiments on multiple unlearning tasks, standard benchmark datasets, and diverse neural architectures, demonstrating that it achieves effective and reliable unlearning while outperforming existing baselines in both accuracy retention and removal fidelity.

Problem

• machine unlearning
• retain-forget entanglement
• feature correlation
• semantic similarity
• unintended forgetting

Innovation

• machine unlearning
• retain-forget entanglement
• augmented Lagrangian
• gradient projection
• Wasserstein-2 regularization