Abstract
Graph Neural Networks (GNNs) are progressively utilized in multiple relevant domains, such as social networks, recommendation systems and cybersecurity. However, their resilience against adversarial attacks has become a significant concern. Existing research has proven that there are adversarial vulnerabilities in GNNs by lightly perturbing the network data to corrupt their performances This study analyzes the transferability of these vulnerabilities across different GNN architectures by employing transfer learning techniques. This approach assesses how adversarial information may spread through distinct models. We begin with a GNN that is trained on adversarial data. This model exhibits specific weaknesses, allowing us to transfer learned representations to another GNN architecture to evaluate how adversarial patterns impact the target model. By training on an additional adversarial dataset, we can assess inherited weaknesses and explore new vulnerabilities unique to the selected architecture. Our research aims to analyze both common and specific susceptibilities among various GNN models, thereby enhancing our understanding of adversarial robustness and informing the development of more effective defenses for applications involving graph-based data protection.