Search for a command to run...
Modern Oracle-based data warehouses are increasingly vulnerable to complex cyberattacks due to their centralized storage of sensitive enterprise data. Traditional intrusion detection systems often lack the semantic granularity and interpretability required to detect evolving threats in structured query environments. This paper presents ORAXAI-Net, a novel explainable AI-driven framework for proactive cybersecurity threat detection in Oracle-based data warehouse systems. The architecture integrates a Temporal-Aware Query-Graph Embedding (TAQGE) module to convert query logs into time-encoded heterogeneous graphs that capture behavioral patterns of user activity. These representations are classified using a Hierarchical Attention-driven Capsule Network (HAC-Net), which preserves structural dependencies and enhances detection accuracy. To ensure transparency, an Integrated Gradient Attention Mapping (IGAM) layer is employed to highlight feature contributions for each prediction. Experimental validation was conducted using the UNSW-NB15 dataset. ORAXAI-Net achieved an accuracy of 95.42%, precision of 94.23%, recall of 95.01%, and an F1-score of 94.61%, outperforming traditional classifiers and explainable baselines such as GCN + LIME and XGBoost + SHAP. The model also demonstrated strong interpretability, achieving an explainability score of 0.92, providing actionable insights for security administrators. Moreover, the framework maintained a low inference time (2.1 ms), indicating suitability for real-time environments. These results affirm the effectiveness of the proposed architecture in delivering both robust threat detection and meaningful explanation within data-intensive Oracle ecosystems.