Graph Attention Networks for Authorization in Hadoop

Hadoop is a widely adopted distributed computing framework that enables large-scale parallel processing through MapReduce and relies on the Hadoop Distributed File System (HDFS) for scalable and fault-tolerant storage. Despite its effectiveness for big data workloads, securing Hadoop environments remains a critical challenge, particularly with respect to authorization, due to system heterogeneity, unresolved vulnerabilities, and both accidental and malicious user behaviors. Traditional policy-based models, such as RBAC and ABAC, provide structured access control but remain inherently static, lacking adaptability and proactive anomaly detection. Behavior-based approaches can partially mitigate these limitations; however, they often introduce additional complexity into the authorization process. By exploiting the adaptive neighborhood selection and weighting capabilities of Graph Attention Networks (GATs), the authorization process in Hadoop can be enhanced by prioritizing the most influential relationships among users, roles, and resources. This capability enables more accurate detection of unauthorized access while alleviating the limitations of static, policy-driven models. Building on this insight, we propose the Authorization-based Heterogeneous Graph Attention Network (AHGAT), a novel learning-based framework that models the authorization landscape as a heterogeneous graph. Its architecture combines a decision network, which identifies the most relevant neighborhood type, with a representation network that refines node embeddings accordingly. To evaluate the performance of AHGAT, experiments were conducted on a custom Hadoop log dataset. The results demonstrate that AHGAT can outperform conventional graph-based learning models, including GAT and GCN, across multiple metrics such as accuracy, precision, F1-score, and ROC-AUC. Furthermore, we assess its adversarial robustness using the Fast Gradient Sign Method (FGSM). The findings show that AHGAT can maintain stable performance under adversarial attacks.