In the content, the authors introduce GNN-Infer, a technique for inferring properties of Graph Neural Networks (GNNs). By converting GNNs to equivalent Feed-forward Neural Networks (FNNs), they propose a method to analyze and generalize behaviors specific to influential structures. The approach involves identifying representative structures, inferring properties on these structures, generalizing them to varying input graphs, and enhancing precision through dynamic feature properties. Experimental results demonstrate the effectiveness of GNN-Infer in defending against backdoor attacks on real-world GNNs.
The content discusses the theoretical foundation behind transforming GNNs into FNNs for property inference. It outlines the steps involved in extracting influential substructures, inferring structure-specific properties, and relaxing structural constraints. The algorithmic process is detailed with examples illustrating each step's application in analyzing and inferring likely properties of popular real-world GNN models.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문