Spatio-Temporal Foundation Model for Network Intrusion Detection Systems (NIDS):

Abstract

Nowadays, the cybersecurity field is facing a significant dilemma caused by the exponential evolution of cyber threat vectors. The emergence of AI-driven intrusion techniques, has rendered traditional detection frameworks increasingly insufficient. Consequently, identifying malicious intent now requires more than simple pattern recognition; it demands a deep understanding of the underlying dynamics of network communication. This thesis proposes a robust Network Intrusion Detection System (NIDS) built upon the complexity of packet flow analogies. By leveraging Hybrid Masked Spatio-Temporal Transformers, we redefine network traffic as a spatio-temporal volume. Unlike classical models, our approach utilizes self-supervised learning to capture the joint evolution of internal packet structures (spatial) and flow sequences (temporal). Through this architecture, the model achieves a profound understanding of both historical attack patterns and ”New” unobserved exploits. By mastering the fundamental ”Grammar” of the network, the proposed system provides not only a solution to detect known attacks through their signatures but also promises a scalable and adaptive resolution for Zero-Day detection.