Abstract:
Graph neural networks (GNNs) have emerged as the state of theart for a variety of graph-related tasks and have been widely com-mercialized in real-world scenarios. Behind its revolutionary repre-sentation capability, the huge training costs also expose GNNs tothe risks of potential model piracy attacks which threaten the intel-lectual property (IP) of GNNs. In this work, we design a novel andeective ownership verication framework for GNNs called GNN-Fingers to safeguard the IP of GNNs. The key design of the proposedframework is two-fold: graph ngerprint construction and robustverication module. With GNNFingers, a GNN model owner canverify if a deployed model is stolen from the source GNN simply byquerying with graph inputs. Besides, GNNFingers could be appliedto various GNN models and graph-related tasks. We extensivelyevaluate the proposed framework on various GNNs designed formultiple graph-related tasks including graph classication, graphmatching, node classication, and link prediction. Our results showthat GNNFingers can robustly distinguish post-processed surrogateGNNs from irrelevant GNNs, e.g., GNNFingers achieves 100% truepositives and 100% true negatives on the test of 200 suspect GNNsof both graph classication and node classication tasks.