PANGA: Attention-based Principal Neighborhood Aggregation for Forecasting Future Cyber Attacks

There has been a significant spike in cyber attacks with serious economic, security, and privacy concerns recently. The effectiveness of older NIDS (Network Intrusion Detection Systems) has been diminishing as cyber attacks become more sophisticated and complex. Majority of research focuses on detecting and classifying attacks, with limited work attempting to forecast the number of cyber attack. Recent works propose using statistical, machine learning, or deep learning models to forecast cyber attacks, but they do not take advantage of the correlation that exists among different devices present in a network. We proposed a model, PANGA that employs a combination of Principal Neighborhood Aggregation(PNA), Gated Recurrent Unit(GRU), and attention layers to effectively capture and leverage spatiotemporal dependencies. It achieves a mean square error of 0.027 and a coefficient of determination of 0.87 on well known CIDDS-001 dataset. Additionally, we performed perturbation analysis by adding gaussian noises to the test data to validate the robustness of the model. The performance of the proposed model remains largely unaffected as long as the standard deviation of the noise was kept below 0.25.