A Robust Watermark-based Fingerprint Framework for GNNs Ownership Verification

To address the issue of model ownership infringement caused by high training costs for Graph Neural Networks (GNNs), the research team proposed a fingerprint framework based on robust watermarks called REMARK. This framework generates carefully designed intra-distribution watermarks to maximize the output differences of GNN models, thereby mitigating performance degradation caused by out-of-distribution (OOD) watermarks. Robust fingerprints are then extracted from these output differences to verify ownership, eliminating the assumption that proxy models must be trained on watermark training sets or expose specific output levels. Extensive experiments show that REMARK achieves advanced accuracy and robustness in ownership verification while maintaining the practicality of protected models.