ETD-FGL: Federated Graph Learning Against Backdoor Attacks Using Topology-aware Defense

Abstract

ETD-FGL: Federated Graph Learning Against Backdoor Attacks Using Topology-Aware Defense This release provides the research implementation of ETD-FGL, a topology-aware defense framework for investigating structural backdoor attacks in federated graph learning. The software includes experimental pipelines for the Cora, PubMed, and Reddit datasets, implementations and research adaptations of baseline methods, trust-weighted aggregation, topology and explanation analysis, and automated generation of publication-quality figures. Key features Evaluation across Cora, PubMed, and Reddit graph datasets. Comparison of clean and attacked federated learning baselines with defense methods. Topology-aware and explanation-based analysis for structural backdoor defense. Extended experiments covering multi-seed evaluation, non-IID robustness, ablation studies, sensitivity analysis, and client scalability. Automated generation of comparison tables and figures in PNG, PDF, and SVG formats. Reproducible experiment configurations and documented execution workflows. Reproducibility and scope The repository provides scripts and configurations for reproducing the supported experiments. Results depend on the selected dataset, experimental configuration, random seed, and computational environment. The Reddit benchmark uses a deterministic sampled subgraph rather than the complete graph. Baseline implementations are research adaptations and should not be interpreted as official implementations unless otherwise specified. Software and documentation Repository: https://github.com/aynazsalamati/ETD-FGL License: MIT Author: Aynaz Salamati This release is intended to support reproducible research, methodological comparison, and further investigation of security and robustness in federated graph learning. For installation instructions, experiment commands, and citation details, please refer to the repository README and CITATION.cff file.