Reproduce “ATOM: A Framework of Detecting Query-Based Model Extraction Attacks for Graph Neural Networks” on OpenResearch

About this paper

Zhan Cheng, Bolin Shen, Tianming Sha, Yuan Gao, Shibo Li, Yushun Dong

The ATOM framework provides a proactive, real-time defense against query-based Model Extraction Attacks (MEAs) targeting Graph Neural Networks (GNNs) in Graph-based Machine Learning as a Service (GMLaaS) environments. This method, integrating graph-aware embeddings and adaptive sequential modeling, consistently outperforms baseline detection methods and demonstrates robust real-time adaptability across various datasets.

Key insights

• Modeling user behavior as a dominating set problem highlights the trade-off between attack efficiency and detection evasion, justifying the integration of k-core values in query embeddings.
• Capturing first and second-order differences in sequential query patterns is critical for detecting subtle strategic shifts by adaptive attackers, which standard sequence models often miss.
• A proactive, adaptive defense that combines graph-specific features with advanced sequential modeling and reinforcement learning can effectively detect complex, evolving query-based MEAs in real-time.

Summary by alphaXiv.