Research Symposium

BIO

Gabriella currently attends Florida State University and is pursuing a Bachelor's degree in Computer Science with a minor in Data Analytics. She's interested in pursuing a career in business intelligence and analytics. During high school, she attended the School for Advanced Studies, where she completed her last two years of high school and obtained an Associate in Arts degree from Miami Dade College. Gabriella has a strong passion for data analytics, and is excited to further develop my skills and knowledge in these fields at FSU.

Evaluating the Effectiveness of Defensive Mechanisms Against Model Extraction Attacks in Graph Neural Networks

Abstract

Model extraction attacks pose a significant threat to the security of machine learning systems by enabling adversaries to replicate deployed models through limited interactions. In graph neural networks (GNNs)—a type of machine learning model designed to learn from data represented as networks of connected nodes, such as social networks or molecular structures—recent advances in explainability have introduced new attack methods by revealing information about a model’s internal reasoning. This project examines the impact of explanation-guided extraction attacks by reproducing a recently proposed framework that aligns surrogate model training with target model explanations. Using PyTorch and torch-geometric, we implement the attack and examine its performance on graph-based datasets. The reproduced results confirm that including explanation alignment substantially increases the effectiveness of model extraction compared to standard query-based approaches. Together, these results establish a strong baseline and motivate future work on defenses that can limit information leakage while maintaining predictive accuracy.

Keywords: AI, Machine learning, Technology, Computer, Training