Abstract
Graph Convolutional Networks (GCNs) are powerful tools for learning from graph data but exhibit significant vulnerability to adversarial structural attacks that manipulate node connections. While various defense strategies focusing independently on the spatial or spectral domains exist, they often fail to leverage the complementary strengths of both perspectives. This paper introduces the Spatial- Spectral Graph Convolutional Network (S²-GCN), a novel framework designed to enhance GCN robustness against structural attacks by synergistically combining spatial and spectral defense mechanisms. S²-GCN comprises two core GCN-based modules operating in parallel. The spectral module utilizes a graph structure derived from learnable low-frequency spectral components, adaptively filtering potential noise. The spatial module employs purified 1-hop and 2-hop neighbor information, using an attention mechanism to dynamically weigh their contributions to node embeddings. These modules share parameters and are trained end-to-end. Comprehensive experiments on benchmark datasets demonstrate that S²-GCN significantly outperforms existing baselines and state-of-the-art defense methods in mitigating the impact of various structural attacks, including Metattack and Nettack.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
"Robust Graph Learning through Spatial-Spectral Synergy against Structural Attacks", Technical Disclosure Commons, (May 04, 2025)
https://www.tdcommons.org/dpubs_series/8077