Scholarlink

<p>Recent approaches that use federated learning to address fairness among different participating federated clients do not mitigate biased model aggregation and variable performance between clients due to data heterogeneity. The study proposes a novel approach that addresses this limitation by combining and observing the training loss distribution from all participating clients. This approach exponentially decays the contribution of the lowest and highest performing clients while assigning higher weights to the clients closer to central value of the distribution. This central value is determined by statistical measures of determining the skew in the distribution during the weighted averaging process. The weights are assigned to the clients by normalizing the Euclidean norm of the scalar difference between each client loss and the calculated central value and ensures fair representation of each contributing client model. The performance of these experiments is evaluated on standard benchmark tabular, textual, and vision datasets, namely NSL-KDD, Huffington Post, the Symptom-Disease dataset from the Flower Datasets library, CIFAR-10, and CIFAR-100 respectively. Dirichlet distribution was used to simulate label heterogeneity by converting IID datasets into non-IID datasets to create heterogeneous data distribution. The study uses 25 and 100 clients ensuring wider coverage of the experiments and presents detailed convergence and variance analyses for each of the training, validation, and testing phases. The variance in client accuracy was approximately 20 times lower than the proposed approach compared to the other methods, without any compromise in client performance in terms of accuracy. This study also presents a detailed comparison of FedDisco and q-FFL state-of-the-art benchmark algorithms that handle fairness in the federate learning. The proposed approach displays improvement in accuracy and uniformity across client accuracy by enhancing the fairness of the federated learning system using a novel distribution-aware algorithm.</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

FedA2W: adaptive aggregation with weighted loss proximity for non-IID federated learning

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Sushant Jain,
  • Sumukh Gupta,
  • Amit Pundir,
  • Sanjeev Singh,
  • Geetika Jain Saxena

摘要

Recent approaches that use federated learning to address fairness among different participating federated clients do not mitigate biased model aggregation and variable performance between clients due to data heterogeneity. The study proposes a novel approach that addresses this limitation by combining and observing the training loss distribution from all participating clients. This approach exponentially decays the contribution of the lowest and highest performing clients while assigning higher weights to the clients closer to central value of the distribution. This central value is determined by statistical measures of determining the skew in the distribution during the weighted averaging process. The weights are assigned to the clients by normalizing the Euclidean norm of the scalar difference between each client loss and the calculated central value and ensures fair representation of each contributing client model. The performance of these experiments is evaluated on standard benchmark tabular, textual, and vision datasets, namely NSL-KDD, Huffington Post, the Symptom-Disease dataset from the Flower Datasets library, CIFAR-10, and CIFAR-100 respectively. Dirichlet distribution was used to simulate label heterogeneity by converting IID datasets into non-IID datasets to create heterogeneous data distribution. The study uses 25 and 100 clients ensuring wider coverage of the experiments and presents detailed convergence and variance analyses for each of the training, validation, and testing phases. The variance in client accuracy was approximately 20 times lower than the proposed approach compared to the other methods, without any compromise in client performance in terms of accuracy. This study also presents a detailed comparison of FedDisco and q-FFL state-of-the-art benchmark algorithms that handle fairness in the federate learning. The proposed approach displays improvement in accuracy and uniformity across client accuracy by enhancing the fairness of the federated learning system using a novel distribution-aware algorithm.