Researchers have developed new methods to test the adversarial robustness of Relational Deep Learning (RDL) pipelines, which are commonly used for machine learning on relational databases. These pipelines encode databases as graphs, with nodes representing tuples and edges representing primary-key to foreign-key dependencies, and then train graph neural networks. The study focuses on a white-box attacker who can rewire foreign-key references in the database while adhering to schema integrity constraints. Seven attack heuristics, including gradient-guided variants, were evaluated on the RelBench rel-f1 benchmark, showing that gradient-based attacks are more effective for regression tasks than for classification. AI
IMPACT Introduces new attack vectors for evaluating the security and robustness of graph neural networks used in database machine learning.
RANK_REASON Academic paper detailing a new methodology for evaluating adversarial attacks on a specific type of machine learning pipeline. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
