HermesHFL: Incentive-Compatible Hierarchical Federated Unlearning for Dynamic LLM Fine-Tuning

2026-07-13

AI Summary

This work addresses the challenge of selective forgetting in large language models within hierarchical federated learning, where layered aggregation, dynamic client participation, and strong parameter coupling hinder effective unlearning. To tackle this, we propose HermesHFL, a novel framework that uniquely integrates incentive-compatible mechanisms with hierarchical federated unlearning. HermesHFL leverages LoRA for parameter-efficient fine-tuning and introduces Neogen, a neural-guided bilevel evolutionary optimizer that jointly optimizes client participation, edge association, incentive allocation, and the forgetting process. By combining CMA-ES and CHC algorithms enhanced with a neural surrogate for acceleration, HermesHFL significantly outperforms existing methods across multiple LLM fine-tuning tasks, achieving state-of-the-art performance in model utility, unlearning efficacy, convergence stability, and resource efficiency.

Abstract

Hierarchical federated unlearning (HFUL) for large language model (LLM) fine-tuning faces significant challenges due to hierarchical aggregation, dynamic client participation, and strong parameter coupling in LLM adaptation. Selectively removing client contributions is particularly difficult because model updates propagate across multiple aggregation stages while unlearning requests may coincide with client departures and rejoining. To address these issues, we propose **HermesHFL**, a hierarchical federated learning framework that supports selective unlearning, dynamic client participation, and client reintegration for scalable LLM fine-tuning via parameter-efficient fine-tuning (PEFT) with LoRA. We formulate a unified optimization problem that jointly models client participation, edge association, incentive allocation, and unlearning under heterogeneous client behaviors. To solve this problem efficiently, we develop **Neogen**, a neural-guided bilevel evolutionary optimization framework that combines CMA-ES for continuous incentive optimization with a CHC-based evolutionary mechanism for discrete participation and association decisions. A neural surrogate further accelerates optimization and improves search efficiency. Extensive experiments on LLM fine-tuning tasks demonstrate that HermesHFL consistently outperforms state-of-the-art baselines in model utility, unlearning effectiveness, convergence stability, and resource efficiency.

Problem

• hierarchical federated unlearning
• large language model
• dynamic client participation
• selective unlearning
• parameter coupling

Innovation

• Hierarchical Federated Unlearning
• Parameter-Efficient Fine-Tuning
• Neural-Guided Bilevel Optimization
• Dynamic Client Participation
• Incentive-Compatible Learning