一种融合数据新鲜度的联邦学习动态激励机制

打开文本图片集

关键词：联邦学习；激励机制；多智能体强化学习；数据新鲜度

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）08-033-2497-06

doi：10.19734/j.issn.1001-3695.2025.01.0012

Dynamic incentive mechanism for federated learning incorporating data freshness

Dawulie Jinensibieke 1，2，3 ，Wang Yi 1，2，3 ， Zhou Xi 1，2，3 ，Wang Xiaobo 1，2，3† （1.XinjiangTclsefi&strecdefec，Uu1Uesitde fSciences，Bino9，ina；.XinjangLbotoryfnoritySpech&LangugeInfomationProcessng，Uruq301，ina）

Abstract：Inmulti-roundfederated learning，thetrainingenvironmentisconstantlychanging.Incentive mechanismsbasedon multi-agentreinforcementlearning（MARL）canbeteradapttodynamicenvironmentsbydynamicallyadjustingdatacontributionstrategies.Existing MARL-basedmechanismsoftenfocusondataquantity，neglectingdatafreshness，whichleadstoicomplete contribution evaluationand limits modelperformance.Thispaper proposed a dynamic incentive mechanismfor federated learning incorporatingdatafreshness.Itintroduceddata freshnessmetricsand benefit evaluationmethods tomorecompreensivelyassesscontributions.Themechanismemployedacentralizedtraining withdecentralizedexecutionMARLframeworkto addresscoordinationissues，maximizingoverallbenefits.Experimentalresultsshowthattheproposedmethodimprovesoveal benefitsby approximately 11.1% to 25.0% across five public datasets.Comparative and ablation experiments further validate the fairness and effectiveness of the method under varying data quality conditions.

KeyWords：federated learning；incentivemechanism；multi-agent reinforcement learning；data freshness

0 引言

联邦学习（federatedlearning，FL）技术允许多个参与方在无须共享原始数据的情况下协同训练模型，有效解决了多方模型训练过程中的隐私和安全问题[1]。（剩余16465字）