knowledge-based Fuzzy Defensive Classi cation Model for Robust MPGNNs

Message-passing graph neural network (MPNN) is a highly e effective learning tool for graph-structured data but is particularly vulnerable to adversarial attacks on information ow, where poisoned data from attacked nodes can propagate to neighboring nodes and be aggregated, amplifying the negative impact across the graph. The malicious perturbations to node features or graph structures can degrade model performance in defensive decision making, which requires a clean ow. We propose a creative defense mechanism to robustly enhance the resilience of MPGNNs against adversarial perturbations by leveraging a creative knowledge-based fuzzy approach for node classification that is implemented by a rule-based fuzzy model. Our study assigns each neighboring node a membership function, which quantifies its relevance to the target node based on knowledge extracted from structural information, node attributes, and community relationships to the target node. By dynamically adjusting the influence of neighboring nodes, the proposed model mitigates the effects of adversarial noise and reduces the impact of irrelevant or poisoned data, ensuring more robust message aggregation. Furthermore, the fuzzy approach helps prevent over-squashing, a common issue when increasing the number of GNN layers, by maintaining important information through controlled message passing and is followed by picking an opinion leader agent from each egonet to recycle the graph. However, our model does not require prior knowledge of the graph or attack types, making it adaptable to various datasets and attack scenarios. Extensive experiments on benchmark datasets, including Cora, PubMed, and Citeseer, demonstrate greater robustness and better accuracy when compared to state-of-the-art defense strategies with reasonable complexity.