AI-Driven Cybersecurity Models for Intelligent Network Protection
DOI:
https://doi.org/10.5281/zenodo.19152792Keywords:
cybersecurity, intrusion detection, graph neural networks, adversarial robustness, network protection, LLM, incident response, AI securityAbstract
The escalating sophistication of cyber threats — encompassing advanced persistent threats, zero-day exploits, AI-generated adversarial attacks, and
large-scale coordinated botnet operations — has rendered signature-based and rule-driven network protection systems fundamentally inadequate.
AI-driven cybersecurity models offer adaptive, behaviour-based threat detection capable of identifying novel attack patterns without pre-existing
signatures, yet their deployment in operational network environments faces challenges of adversarial evasion, explainability for incident response,
and real-time throughput under high-bandwidth network conditions. This study presents CyberAI, a unified AI-driven cybersecurity framework for
intelligent network protection integrating a graph neural network-based network behaviour analyser, an adversarially robust ensemble intrusion
detection system, and a large language model (LLM)-assisted incident response recommendation engine. CyberAI was evaluated on the
CIC-IDS-2018, UNSW-NB15, and CAIDA DDoS benchmarks, and deployed on the Chalmers University enterprise network (12,400 active endpoints,
40 Gbps backbone) and the UC3M campus network (8,700 endpoints, 20 Gbps). CyberAI achieved intrusion detection F1-score of 0.984 on
CIC-IDS-2018, adversarial attack detection accuracy of 94.7% under PGD perturbation, mean detection latency of 12.3 ms, and incident response
recommendation precision of 0.931. False positive rate of 0.3% in real deployment represents a 74.1% reduction from the incumbent SIEM baseline.
These results establish CyberAI as an accurate, adversarially robust, and operationally deployable AI-driven cybersecurity framework for intelligent
network protection.
Keywords: ;; ; ; ; ;;
