.png){kind=icon}
Return
A Distributed Difference-Aware Alignment Framework for Robust Federated Graph Learning
.png){kind=icon}
PRE
AI
DOI：10.1109/tnse.2026.3718246.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Federated Graph Learning (FGL) has emerged as a crucial paradigm for collaborative training across distributed graph data. However, in Non-Independent and Identically Distributed (non-IID) scenarios, heterogeneous data partitions and representational distortion introduced by graph augmentations lead to node-level semantic drift, which restricts the generalization of the global model. Existing approaches mainly rely on graph augmentation or aggregation strategies, but often assume that augmentations preserve semantics, overlooking the drift between original and augmented graphs and leading to local model instability. To address this, we propose Federated Difference-aware Semantic-alignment Learning (FedDSL), a framework that jointly tackles augmentation-induced and distribution-induced drift. Specifically, we design a difference-aware regularization module that quantifies attribute and structural differences between original and augmented graphs to stabilize local embeddings with modest additional computation. Furthermore, a node-level semantic alignment module is proposed to align local embeddings with global representations, improving cross-client consistency without introducing extra communication rounds. Extensive experiments on six benchmark datasets, covering both homophilic and heterophilic graphs, demonstrate that FedDSL consistently outperforms state-of-the-art baselines under non-IID conditions.
Keywords:
Federated graph learning
semantic drift mitigation
distributed representation alignment
non-IID data
cross-client consistency
robust distributed training
Journal
I
IF:
7.9
Papers:
2.6K
Citations:
10.0K
Organization
Cited Papers
No cited papers available