Hierarchical Federated Unlearning for Large Language Models

2025-10-19

AI Summary

To address inter- and intra-domain interference and the trade-off between forgetting efficacy and task retention in decentralized continual unlearning for large language models (LLMs), this paper proposes a Hierarchical Federated Unlearning (HFU) framework. Built upon federated learning, HFU employs task-specific adapters to decouple parameter updates for forgetting and retention objectives. A hierarchical model fusion strategy—operating jointly at client and server levels—mitigates interference arising from data heterogeneity and asymmetric access. The method enables efficient, scalable, privacy-preserving model updates without centralized data collection. Evaluated on the WMDP, MUSE, and TOFU benchmarks, HFU significantly outperforms existing baselines: it achieves strong, targeted unlearning while preserving downstream utility of LLMs to the greatest extent possible.

Technology Category

• Machine Learning: Distributed Machine Learning & Federated Learning
• Natural Language Processing: (Large) Language Models
• Humans and AI: Human-in-the-loop Machine Learning

Application Category

• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

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.

Problem

• Removing undesirable knowledge from large language models
• Handling continuous heterogeneous unlearning requests in decentralized data
• Mitigating performance conflicts between forgetting and retaining objectives

Innovation

• Hierarchical merging strategy mitigates conflicting objectives
• Task-specific adapter learning decouples unlearning and retention
• Federated approach handles heterogeneous unlearning requests scalably