FedSal: Enhancing Federated Graph Classification Through Saliency Aware Client Clustering

ICLR 2026 Conference Submission22644 Authors

Keywords: Federated Learning, Graph Neural Network, Saliency Maps, Federated Graph Neural networks

TL;DR: ChatGPT said: FedSal is the first federated GNN using client-side saliency maps for adaptive clustering under severe non-IID heterogeneity. In experiments it consistently outperforms SOTA FedGNN baselines in graph classification.

Abstract: Graph Neural Networks (GNNs) are essential for analyzing structured data but face significant challenges in federated learning (FL) environments, where non-IID client distributions and structural heterogeneity impede convergence and performance. To address these issues, we introduce Federated Saliency Aggregation Learning (FedSal), the first framework to apply saliency maps in GNN-based FL on graph classification tasks. FedSal replaces full-gradient uploads with compact saliency activations, enabling dynamic clustering of clients via simple thresholds (\(\epsilon_{\mathrm{mean}}, \epsilon_{\max}\)) and cluster-wise model averaging. We further propose FedSal+, which augments node features with positional and random-walk encodings to inject structural priors without exposing raw graph data. Extensive experiments on thirteen molecular, protein, and social-network benchmarks under extreme non-IID splits show that FedSal and FedSal+ achieve higher accuracy, converge faster, and reduce communication cost compared to state-of-the-art methods. These results demonstrate the SOTA performance of saliency-driven clustering for personalized, robust, and communication-efficient federated graph classification tasks.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 22644