中国空间科学技术

中国空间科学技术 ›› 2026, Vol. 46 ›› Issue (5): 66-74.doi: 10.16708/j.cnki.1000-758X.2026.0077

• 数学与航天专题 • 上一篇    下一篇

DSM：基于领域漂移的分布节点间Non-IIDness的度量方法

李澔1，刘雪娇2，陈千喜1，李岳1,*   

  1. 1.南开大学计算机学院，天津300350
    2.中国空间技术研究院钱学森空间技术实验室，北京100094
  • 收稿日期:2025-10-27 修回日期:2026-01-27 录用日期:2026-03-25 发布日期:2026-07-31 出版日期:2026-09-30

DSM: a domain-shift-based metric for measuring Non-IIDness across distributed nodes

LI Hao1，LIU Xuejiao2，CHEN Qianxi1，LI Yue1,*   

  1. 1.College of Computer Science，Nankai University，Tianjin 300350, China
    2.Qian Xuesen Laboratory of Space Technology, China Academy of Space Technology，Beijing 100094，China
  • Received:2025-10-27 Revision received:2026-01-27 Accepted:2026-03-25 Online:2026-07-31 Published:2026-09-30

PDF (PC)

34

摘要： 联邦学习面临的核心挑战之一，是如何有效解决客户端数据的非独立同分布特异性。现有研究缺乏一种能够跨越不同非独立同分布场景、统一量化数据差异并公平评估算法鲁棒性的通用度量指标。本研究旨在解决这一空白，探索对分布式节点间非独立同分布程度的普适量化方法。本研究提出一种基于领域漂移的度量方法（DSM）。该方法的核心思想是，通过预训练的特征提取网络，将难以直接计算的高维原始数据分布差异，映射到特征空间中，转化为可计算的领域漂移问题。具体而言，通过计算各节点特征分布与全局特征分布之间的Wasserstein距离，并进行加权聚合，从而实现对非独立同分布程度的、与模型无关的定量描述。在MNIST和Fashion-MNIST数据集上，构建了5类典型的非独立同分布场景进行验证。实验结果表明：DSM能够有效区分这5类场景内部及相互之间的偏置程度，其灵敏度显著优于欧氏距离、余弦相似度等4种传统非参数度量方法；利用DSM将不同场景校准至相同的非独立同分布程度后，对4种联邦分类算法和4种联邦生成算法进行了公平评估。评估一致表明，在分类任务中Scaffold算法综合性能最优，在生成任务中BGAN算法表现最佳。这证明了DSM支持跨场景公平比较的有效性。本工作提出的DSM度量，首次为联邦学习提供了跨非独立同分布场景的统一量化基准。该方法是模型无关的，不仅为算法评估与优选提供了理论工具，也为应对航天卫星星座等天然具有非独立同分布特性的分布式系统协同学习挑战，提供了关键的量化设计依据。

关键词: 联邦学习, 非独立同分布, 分布差异度量, 领域适应, 星载分布式学习

Abstract: Non-independent and identically distributed (Non-IID) data across clients remains a fundamental challenge in federated learning. A critical gap exists: the lack of a universal metric capable of both quantifying distributional differences across diverse Non-IID scenarios and fairly evaluating algorithmic robustness. This work addresses this gap by developing a generic metric for the degree of Non-IIDness among distributed nodes. We propose a Domain-shift-based Metric (DSM). The core idea is to circumvent the intractability of measuring divergence in high-dimensional raw data by leveraging a pre-trained feature extractor. DSM transforms the problem into estimating domain shift within a learned feature space. Specifically, it quantifies the Non-IID degree as theweighted sum of Wasserstein distances between client feature distributions and the global feature distribution. We validated DSM on MNIST and Fashion-MNIST datasets, simulating five canonical Non-IID scenarios. Our experiments demonstrate that: DSM effectively differentiates bias levels both within and across these five scenarios, showing superior sensitivity compared to four conventional metrics (e.g., Euclidean, Cosine distances); After calibrating different scenarios to an equivalent Non-IID degree using DSM, we conducted a fair comparison of four federated classification algorithms and four federated generative algorithms. The evaluation consistently identified Scaffold as the top performer for classification and BGAN for generation, validating DSM's utility for fair cross-scenario algorithm assessment. The proposed DSM metric establishes the first unified, quantitative benchmark for Non-IIDness across scenarios in federated learning. As a model-agnostic tool, it provides a principled basis for algorithm evaluation and selection. Furthermore, it offers a crucial quantitative foundation for designing collaborative learning algorithms in distributed systems with inherent Non-IID data, such as satellite constellations in aerospace applications.

Key words: federated learning, Non-IID, distribution discrepancy measurement, domain adaptation, spaceborne distributed learning