<p>Graph Neural Networks (GNNs) have recently achieved remarkable progress across various domains. In distributed machine learning systems, the inherent communication and collaboration among nodes naturally forms a graph structure, providing an ideal data foundation for GNN applications. Leveraging this intrinsic representation, GNN techniques enable not only efficient processing of complex graph data but also the detection and analysis of anomalous communication patterns and malicious node behaviors, thereby serving as an early warning mechanism for system security. However, despite their excellent performance, state-of-the-art GNN models remain vulnerable to adversarial perturbations; even minor modifications can lead to significant performance degradation, particularly through targeted attacks that exploit the message-passing mechanism to manipulate node representations and undermine overall classification accuracy. To address these challenges, we propose SpiderWeb Protector, an efficient defense mechanism inspired by the adaptive repair and selective reinforcement of spider webs. By harnessing structural graph knowledge to enhance feature representations and dynamically regulating inter-layer information propagation, SpiderWeb Protector comprises three key modules: a community-centric graph update module that reinforces critical graph structures, a topology feature merging module that integrates structural attributes to refine node features, and an edge pruning module that dynamically removes adversarial edges. Experimental evaluations across four GNN models, four defense strategies, and four benchmark datasets demonstrate that SpiderWeb Protector not only outperforms existing methods in mitigating targeted attacks but also provides an effective means to detect and counteract abnormal communication and malicious behaviors in distributed machine learning systems, thereby exhibiting significant potential for practical applications.</p>