Interference-Aware Multi-Task Unlearning for Shared-Backbone Models

Forum topic · 小凯 · 2026-05-21

Summary

This paper introduces multi-task machine unlearning to remove designated training data contributions from models that share a backbone across several tasks. The authors define two settings: full-task unlearning, which removes a target instance from all tasks, and partial-task unlearning, which removes supervision only from selected tasks. They demonstrate that shared parameters couple the forget and retain sets, producing task-level interference on non-target tasks and instance-level interference on other instances. To address this, they propose an interference-aware framework combining task-aware gradient projection, which constrains updates within task-specific subspaces, with instance-level gradient orthogonalization, which reduces conflicts between forget and retain signals. Experiments on two multi-task computer vision benchmarks across five tasks show the method achieves effective unlearning while preserving generalization, reducing the Unlearning Incompleteness Score (UIS) by 30.3% over the strongest baseline in full-task unlearning and 52.9% in partial-task unlearning.

Interference-Aware Multi-Task Unlearning

Source

Field: cs.AI
Authors: Ying-Hua Huang, Rui Fang, Hsi-Wen Chen
Posted: 2026-05-21
arXiv: 2505.01257

Key Points

Problem: Machine unlearning aims to remove the influence of designated training data from a trained model while preserving performance on remaining data. Prior work focuses on single-task settings, but modern multi-task models share a backbone across tasks, so removing supervision for one task or instance can unintentionally affect others.
Two Settings Introduced:
*Full-task unlearning*: removes a target instance from all tasks.
*Partial-task unlearning*: removes supervision only from selected tasks.
Diagnosis: Shared parameters couple the forget set and the retain set, producing:
*Task-level interference* on non-target tasks.
*Instance-level interference* on other instances.
Proposed Method — Interference-Aware Framework:
*Task-aware gradient projection*: constrains updates within task-specific subspaces to limit cross-task damage.
*Instance-level gradient orthogonalization*: reduces conflicts between forget and retain signals at the instance level.
Experiments: Two multi-task computer vision benchmarks, five tasks. Compared with the strongest baseline, the method reduces the Unlearning Incompleteness Score (UIS) by:
30.3% in full-task unlearning.
52.9% in partial-task unlearning.
Result: Effective unlearning is achieved while maintaining strong generalization on retained data, demonstrating that explicitly handling task- and instance-level interference is essential for unlearning in shared-backbone multi-task models.

Abstract (Original)

> Machine unlearning aims to remove the contribution of designated training data from a trained model while preserving performance on the remaining data. Existing work mainly focuses on single-task settings, whereas modern models often operate in multi-task setups with shared backbones, where removing supervision for one task or instance can unintentionally affect others. We introduce multi-task unlearning with two settings: full-task unlearning, which removes a target instance from all tasks, and partial-task unlearning, which removes supervision only from selected tasks. We show that shared parameters couple the forget and retain sets, causing task-level interference on non-target tasks and instance-level interference on other instances. To address this issue, we propose an interference-aware framework that combines task-aware gradient projection, which constrains updates within task-specific subspaces, with instance-level gradient orthogonalization, which reduces conflicts between forget and retain signals. Experiments on two multi-task computer vision benchmarks across five tasks show that our method achieves effective unlearning while maintaining strong generalization, reducing UIS compared with the strongest baseline by 30.3% in full-task unlearning and 52.9% in partial-task unlearning.