CopyCop Algorithm Detects Copycat Graph Neural Networks
A new algorithm called CopyCop can determine whether two Graph Neural Networks (GNNs) were trained independently or if one was designed to mimic the other. The method works even when the GNNs have different architectures, weights, and embedding dimensions, and when the adversarial GNN transforms its output embeddings. CopyCop outperforms existing watermarking and fingerprinting techniques, offering theoretical guarantees. Experiments on 14 datasets and 5 GNN architectures show high accuracy and robustness against various adversarial attacks and transformations. The code is publicly available.
Key facts
- CopyCop identifies copycat GNNs despite different architectures, weights, and embedding dimensions.
- Adversarial GNNs may transform output embeddings to obscure copying.
- Existing watermarking and fingerprinting methods fail under these conditions.
- CopyCop provides theoretical guarantees for ownership verification.
- Tested on 14 datasets and 5 GNN architectures.
- Robust against a broad class of adversarial attacks and transformations.
- Code is available at the provided URL.
- Published on arXiv under Computer Science > Machine Learning.
Entities
Institutions
- arXiv