Scholarlink

<p>Federated learning (FL) is a paradigm for training deep neural network (DNN) models on a distributed set of client devices without sharing local datasets. An FL server updates the global model by aggregating client models’ weights and transferring the updated model back to clients for future training rounds. One key aspect of FL is the disparity in data quality across distributed clients. As a result, the data used for model training on different devices is non-independent and identically distributed (non-IID) and suffers from quantity, label, and feature skewnesses. Data distribution and quality heterogeneities introduce biases that prevent the global model from achieving the desired convergence. Therefore, there is a need to select clients with relatively more balanced and better-quality data. This paper proposes <i>I2Q-FL</i>, a novel heuristic-based client selection mechanism for FL with heterogeneous data distribution. We formulate a metric called the <i>IID-index</i> that uses feature information and enables the FL server to rank the clients. Next, we propose a scheduling algorithm to engage clients in the model training process, ensuring the participation count is restricted within an upper bound. We conducted an extensive empirical study by implementing <i>I2Q-FL</i> (<a href="https://github.com/PriyankaDas-16/I2Q-FL">https://github.com/PriyankaDas-16/I2Q-FL</a>) on an open-source FL evaluation platform called <i>FedEval</i>. We compared our approach with seven baselines, including the centralized model training approach on four publicly available and one real-life dataset for grayscale and RGB images. Experimental results demonstrate that <i>I2Q-FL</i> is scalable and achieves convergence in a reasonable time. It selects clients with balanced and better quality data, resulting in model accuracy marginally lower than the centralized approach, but outperforming the others. Except the centralized model, <i>I2Q-FL</i> outperforms all baseline by 4.1% on an average across all the five baselines.</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

I2Q-FL: an image informativeness quotient-based client selection mechanism for federated learning

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Priyanka Das,
  • Nirnay Ghosh,
  • Soumajit Pramanik,
  • Subhajit Siddhanta

摘要

Federated learning (FL) is a paradigm for training deep neural network (DNN) models on a distributed set of client devices without sharing local datasets. An FL server updates the global model by aggregating client models’ weights and transferring the updated model back to clients for future training rounds. One key aspect of FL is the disparity in data quality across distributed clients. As a result, the data used for model training on different devices is non-independent and identically distributed (non-IID) and suffers from quantity, label, and feature skewnesses. Data distribution and quality heterogeneities introduce biases that prevent the global model from achieving the desired convergence. Therefore, there is a need to select clients with relatively more balanced and better-quality data. This paper proposes I2Q-FL, a novel heuristic-based client selection mechanism for FL with heterogeneous data distribution. We formulate a metric called the IID-index that uses feature information and enables the FL server to rank the clients. Next, we propose a scheduling algorithm to engage clients in the model training process, ensuring the participation count is restricted within an upper bound. We conducted an extensive empirical study by implementing I2Q-FL (https://github.com/PriyankaDas-16/I2Q-FL) on an open-source FL evaluation platform called FedEval. We compared our approach with seven baselines, including the centralized model training approach on four publicly available and one real-life dataset for grayscale and RGB images. Experimental results demonstrate that I2Q-FL is scalable and achieves convergence in a reasonable time. It selects clients with balanced and better quality data, resulting in model accuracy marginally lower than the centralized approach, but outperforming the others. Except the centralized model, I2Q-FL outperforms all baseline by 4.1% on an average across all the five baselines.