ADAGE: Active Defenses Against GNN Extraction
Researchers have introduced ADAGE, a novel active defense system designed to prevent the theft of Graph Neural Networks (GNNs). Unlike previous defenses that focused on identifying stolen models, ADAGE proactively monitors query diversity and perturbs outputs to make model extraction infeasible. Experiments demonstrate that ADAGE effectively deters attackers across various GNN models and datasets while maintaining high predictive performance for legitimate users. AI
IMPACT This research introduces a novel defense against model extraction for GNNs, potentially enhancing the security of AI models in sensitive applications like drug discovery and traffic prediction.