Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hierarchical Federated Unlearning for Large Language Models

Yisheng Zhong, Zhengbang Yang, Zhuangdi Zhu

TL;DR

This work addresses the challenge of removing undesirable knowledge from large language models in a privacy-preserving, scalable manner when unlearning requests are decentralized and heterogeneous. It proposes Federated UnLearning Merge (FULM), a framework that decouples unlearning and retention via task-specific adapters and employs hierarchical merging to mitigate inter- and intra-domain interference. The method uses LoRA adapters and a two-stage merging process—intra-cluster voting-based merging (TIES) and inter-cluster additive merging—guided by cosine similarity among task vectors. Experiments on WMDP, TOFU, and MUSE demonstrate that FULM effectively forgets targeted content while preserving utility, outperforming baselines in both near-iid and heterogeneous settings. This yields a practical, privacy-preserving approach to LLM knowledge governance with strong scalability for continual unlearning tasks, without centralized data sharing.

Abstract

Large Language Models (LLMs) are increasingly integrated into real-world applications, raising concerns about privacy, security and the need to remove undesirable knowledge. Machine Unlearning has emerged as a promising solution, yet faces two key challenges: (1) practical unlearning needs are often continuous and heterogeneous, and (2) they involve decentralized, sensitive data with asymmetric access. These factors result in inter-domain and intra-domain interference, which further amplifies the dilemma of unbalanced forgetting and retaining performance. In response, we propose a federated unlearning approach for LLMs that is scalable and privacy preserving. Our method decouples unlearning and retention via task-specific adapter learning and employs a hierarchical merging strategy to mitigate conflicting objectives and enables robust, adaptable unlearning updates. Comprehensive experiments on benchmarks of WMDP, MUSE, and TOFU showed that our approach effectively handles heterogeneous unlearning requests while maintaining strong LLM utility compared with baseline methods.

Hierarchical Federated Unlearning for Large Language Models

TL;DR

This work addresses the challenge of removing undesirable knowledge from large language models in a privacy-preserving, scalable manner when unlearning requests are decentralized and heterogeneous. It proposes Federated UnLearning Merge (FULM), a framework that decouples unlearning and retention via task-specific adapters and employs hierarchical merging to mitigate inter- and intra-domain interference. The method uses LoRA adapters and a two-stage merging process—intra-cluster voting-based merging (TIES) and inter-cluster additive merging—guided by cosine similarity among task vectors. Experiments on WMDP, TOFU, and MUSE demonstrate that FULM effectively forgets targeted content while preserving utility, outperforming baselines in both near-iid and heterogeneous settings. This yields a practical, privacy-preserving approach to LLM knowledge governance with strong scalability for continual unlearning tasks, without centralized data sharing.

Abstract

Large Language Models (LLMs) are increasingly integrated into real-world applications, raising concerns about privacy, security and the need to remove undesirable knowledge. Machine Unlearning has emerged as a promising solution, yet faces two key challenges: (1) practical unlearning needs are often continuous and heterogeneous, and (2) they involve decentralized, sensitive data with asymmetric access. These factors result in inter-domain and intra-domain interference, which further amplifies the dilemma of unbalanced forgetting and retaining performance. In response, we propose a federated unlearning approach for LLMs that is scalable and privacy preserving. Our method decouples unlearning and retention via task-specific adapter learning and employs a hierarchical merging strategy to mitigate conflicting objectives and enables robust, adaptable unlearning updates. Comprehensive experiments on benchmarks of WMDP, MUSE, and TOFU showed that our approach effectively handles heterogeneous unlearning requests while maintaining strong LLM utility compared with baseline methods.
Paper Structure (19 sections, 4 equations, 2 figures, 13 tables, 1 algorithm)

This paper contains 19 sections, 4 equations, 2 figures, 13 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Method
  5. Split-and-Merge: Decoupled Unlearning and Retenion via Dual Task adapters
  6. Multi-Task Unlearning Transfer via Hierarchical Federated Merging
  7. Step 1: Intra-Cluster Merging.
  8. Step 2: Inter-Cluster Merging.
  9. Experiments
  10. Experimental Setup
  11. Performance Evaluation
  12. Effects of Hierarchical Merging:
  13. Effects of Decoupled Unlearning and Retention :
  14. Breaking Down Analysis of Hierarchical Unlearning Merging
  15. Conclusion
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Our method tackles heterogeneous, decentralized LLM unlearning requests without transmitting the sensitive unlearning or retention data.
  • Figure 2: Cosine similarity across task vectors, where vectors from near-iid sources exhibit high similarity, and those from heterogeneous domains are nearly orthogonal. Task vectors trained for retention ($TOFU_r$) show negative correlation to their unlearning counterparts ($TOFU$).