Scholarlink

<p>Federated learning (FL) confronts a fundamental paradox: conventional aggregation strategies enforce consensus that stifles the diversity making federated systems powerful. This consensus-driven approach degrades the performance in non-IID environments, where conflicting client updates drive the global model into suboptimal local minima. We introduce Federated Diversity Exploration (<span>FedDive</span>), which rewards divergence instead of penalizing it. <span>FedDive</span> maintains a momentum-guided trajectory representing global consensus and then amplifies the influence of clients whose updates diverge from this path. This exploration-rewarding mechanism enables escape from shallow local minima, transforming client diversity into an optimization asset. This claim is validated on the MNIST dataset, given the model architecture’s and the hyperparameters’ dependence on the dataset, with extensions to more complex data environments identified as a key direction for future work. Experiments demonstrate <span>FedDive</span>’s effectiveness: maintaining performance on IID data at 97.78% (versus <span>FedAvg</span>’s 97.92%), achieving 88.50% accuracy under extreme heterogeneity (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\alpha =0.1\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>α</mi> <mo>=</mo> <mn>0.1</mn> </mrow> </math></EquationSource> </InlineEquation>) (versus <span>FedAvg</span>’s 79.62%), and exhibiting up to 2.67x faster convergence under pathological non-IID conditions, identifying optimal temperature settings for exploration–exploitation balance, and introducing <span>FedDive-R</span>, a robust variant with gated divergence rewards that prevent malicious clients from exploiting the reward mechanism, thereby maintaining 97.01% accuracy even under outlier attacks. The approach reimagines federated aggregation by harnessing divergent clients’ exploratory power while distinguishing beneficial exploration from malicious attacks, proving that properly gated diversity is key to transcending consensus limitations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FedDive: escaping local minima through divergence-rewarded exploration in federated learning: a robust approach with gated divergence rewards

  • 期刊论文

  • 发表日期：

    2025年12月30日

查看全文
  • Dhanraj Rateria,
  • G. M. Siddesh,
  • Lakshay Laddha,
  • Divya Prakash,
  • Pratham Singh,
  • S. R. Mani Sekhar

摘要

Federated learning (FL) confronts a fundamental paradox: conventional aggregation strategies enforce consensus that stifles the diversity making federated systems powerful. This consensus-driven approach degrades the performance in non-IID environments, where conflicting client updates drive the global model into suboptimal local minima. We introduce Federated Diversity Exploration (FedDive), which rewards divergence instead of penalizing it. FedDive maintains a momentum-guided trajectory representing global consensus and then amplifies the influence of clients whose updates diverge from this path. This exploration-rewarding mechanism enables escape from shallow local minima, transforming client diversity into an optimization asset. This claim is validated on the MNIST dataset, given the model architecture’s and the hyperparameters’ dependence on the dataset, with extensions to more complex data environments identified as a key direction for future work. Experiments demonstrate FedDive’s effectiveness: maintaining performance on IID data at 97.78% (versus FedAvg’s 97.92%), achieving 88.50% accuracy under extreme heterogeneity ( \(\alpha =0.1\) α = 0.1 ) (versus FedAvg’s 79.62%), and exhibiting up to 2.67x faster convergence under pathological non-IID conditions, identifying optimal temperature settings for exploration–exploitation balance, and introducing FedDive-R, a robust variant with gated divergence rewards that prevent malicious clients from exploiting the reward mechanism, thereby maintaining 97.01% accuracy even under outlier attacks. The approach reimagines federated aggregation by harnessing divergent clients’ exploratory power while distinguishing beneficial exploration from malicious attacks, proving that properly gated diversity is key to transcending consensus limitations.