Multi-Objective Reference-Aligned Machine Unlearning

2026-05-29

AI Summary

This work addresses the challenges of catastrophic forgetting and degraded model utility in machine unlearning caused by single-objective optimization. To this end, the authors propose RAUL, a novel framework that formulates unlearning as a multi-objective optimization problem for the first time. RAUL simultaneously achieves precise sample removal by aligning with a reference distribution under a bounded KL divergence constraint and preserves pre-trained knowledge. By leveraging the reference distribution to constrain the unlearning objective and integrating multi-objective optimization with Jacobian descent to harmonize conflicting gradients, RAUL effectively mitigates gradient interference. Experimental results demonstrate that RAUL achieves the best trade-off between unlearning efficacy and model utility, yielding performance closest to full retraining.

Abstract

Machine unlearning aims to remove the influence of specific training samples while preserving the model's utility. Existing single-objective approaches, such as gradient ascent or random relabeling, often induce catastrophic forgetting due to conflicting optimization dynamics and unbounded forgetting objectives that cause the model to drift from its pre-trained knowledge. We propose Reference-Aligned UnLearning (RAUL), a multi-objective framework that jointly optimizes forgetting and retention by replacing unbounded loss maximization with a bounded KL alignment of predictions on forgotten samples toward a reference distribution representing unseen data, instantiated either as a uniform distribution or an empirical distribution from a held-out reference set, which constrains the forgetting objective and reduces gradient conflict with retention. The resulting multi-objective optimization (MOO) problem is solved via Jacobian descent, which aggregates multiple gradients into a direction that does not conflict. Our results demonstrate that RAUL achieves the closest gap compared to full retraining.

Problem

• machine unlearning
• catastrophic forgetting
• multi-objective optimization
• gradient conflict
• model utility

Innovation

• machine unlearning
• multi-objective optimization
• KL alignment
• reference distribution
• Jacobian descent