异构联邦持续学习的自适应多因素遗忘方法
Adaptive Multi-factor Forgetting for Heterogeneous Federated Continual Learning
余澄昊 1常晓林1
作者信息
- 1. 北京交通大学网络空间安全学院，北京市 100044
- 折叠
摘要
联邦持续学习旨在隐私约束下实现多客户端的连续任务学习，但在高度异构环境下，客户端间数据分布与任务差异显著，传统方法在应对灾难性遗忘的同时，往往忽视负迁移问题，导致无关甚至有害知识被持续保留，进而限制模型性能提升。针对这一挑战，本文从"完全记忆"向"选择性遗忘"转变，提出一种面向异构联邦持续学习的自适应多因素遗忘方法 AMAF-FCL。该方法首先在特征空间构建归一化流生成模型实现历史知识重构；其次设计多因素样本可靠性评估机制，从三个维度联合量化生成样本质量；在此基础上提出异构感知动态遗忘策略，根据客户端分布差异自适应调节遗忘强度。在多个基准数据集上的实验结果表明，AMAF-FCL在平均准确率与遗忘率指标上均优于现有主流方法，尤其在高异构及含噪场景下表现出更强的鲁棒性。
Abstract
Federated Continual Learning (FCL) aims to achieve sequential task learning across multiple clients under privacy constraints. However, in highly heterogeneous environments, there exist significant discrepancies in data distributions and task objectives among clients. While traditional methods address catastrophic forgetting, they often overlook the negative transfer problem, leading to the persistent retention of irrelevant or even harmful knowledge, which further limits the improvement of model performance.To tackle this challenge, this paper shifts the research paradigm from "complete memorization" to "selective forgetting" and proposes an Adaptive Multi-factor Forgetting method for heterogeneous federated continual learning, namely AMAF-FCL. Specifically, this method first constructs a normalizing flow generative model in the feature space to realize historical knowledge reconstruction. Secondly, a multi-factor sample reliability evaluation mechanism is designed to jointly quantify the quality of generated samples from three dimensions. On this basis, a heterogeneity-aware dynamic forgetting strategy is proposed to adaptively adjust the forgetting intensity according to the distribution differences among clients.Experimental results on multiple benchmark datasets demonstrate that AMAF-FCL outperforms existing state-of-the-art methods in terms of average accuracy and forgetting rate, and exhibits stronger robustness especially in highly heterogeneous and noisy scenarios.关键词
联邦持续学习/选择性遗忘/生成回放/环境异构Key words
Federated Continual Learning/Selective Forgetting/Generative Replay/Environmental Heterogeneity引用本文复制引用
余澄昊,常晓林.异构联邦持续学习的自适应多因素遗忘方法[EB/OL].(2026-05-14)[2026-05-16].http://www.paper.edu.cn/releasepaper/content/202605-49.学科分类
计算技术、计算机技术
评论