DCCF: Dynamic Co-Optimizing Client Selection and Gradient Compression for Efficient-Robust FL under Non-IID Data

Xuehua Sun

doi:10.1016/j.eswa.2026.132239

DCCF: Dynamic Co-Optimizing Client Selection and Gradient Compression for Efficient-Robust FL under Non-IID Data

2026-03-30

0

PRE

AI

DOI：10.1016/j.eswa.2026.132239

Original

Original request for help

Save

Abstract

En 中文

Federated Learning (FL) enables decentralized collaborative training method without compromising privacy of participants. However, in Non-Independent and Identically Distributed(Non-IID) data scenarios, it often encounters issues including delayed convergence, unstable performance, and high communication costs. Existing approaches face the following limitations: (1) client selection mechanisms often rely on predetermined trade-off strategies that lack adaptability to dynamic shifts in client contributions during training; and (2) many mainstream frameworks either adopt a sequential ”select-then-compress” paradigm, or rely on system performance metrics (e.g., computation latency, channel gain). Such designs struggle to client selection that is inconsistent with the gradient compression objective, making it difficult to balance performance and efficiency under data heterogeneity. This paper proposes a Discrepancy-Metric based Client Selection and Compression Framework (DCCF), its core innovation resides in dynamically coordinating client selection and personalized compression through a unified discrepancy metric, achieving dynamic synergistic optimization of both. Specifically, the framework first synchronously assesses the consistency between the local data distribution of each client and the global model via this discrepancy metric, then adaptively selects participating clients based on the evaluation results. This approach balances model consistency and data diversity, thereby suppressing the drift of local models. Furthermore, personalized compression ratios are assigned to each selected client using the discrepancy metric: clients with low discrepancy values are allocated high compression ratios to conserve communication bandwidth, while those with high discrepancy are assigned low compression ratios to preserve crucial gradient information. Additionally, the framework integrates a residual accumulation mechanism to alleviate compression-induced errors and enhance the stability of the training process. More importantly, instead of naively combining client selection and gradient compression, we highlight that their decisions are tightly coupled. Both theoretical analyses and experimental validations demonstrate the necessity of such joint optimization. Experimental results indicate that DCCF outperforms existing mainstream methods in terms of model accuracy, convergence speed, and communication efficiency. Particularly under the extremely Non-IID scenario of the CIFAR-10 dataset, DCCF not only reduces the communication overhead drastically to 7.9% of the baseline method but also accelerates the model convergence rate by 56%.

Keywords:

Federated Learning

Non-IID Data

Client Selection

Gradient Compression

Discrepancy Metric

DCCF: Dynamic Co-Optimizing Client Selection and Gradient Compression for Efficient-Robust FL under Non-IID Data

Abstract

Journal

Organization