Research on AI-Driven Intelligent Perception of Cybersecurity Threats and Dynamic Defense Mechanisms
Fenglin Guan, Hui Ye, Junyi Zeng, Jinhang Liu, Chunzhi Li, Weilong Wen · 2025
In view of the disconcerting escalation in cyberattacks facilitated by artificial intelligence, which presently constitute 67% of such incidences (CISA, 2023) [1], this study proffers a cybersecurity framework that integrates spatiotemporal feature modelling and dynamic decision optimisation. Conventional rule-based defence systems are limited by their one-dimensional perception mechanisms [2], resulting in an average response delay of over 20 minutes. This delay has a detrimental effect on their effectiveness in addressing sophisticated attack scenarios.The proposed architecture demonstrates significant advantages when simulating 30 types of attacks by constructing an LSTM-GCN dual-stream feature extraction network and a game theory-driven reinforcement learning decision-making module. The threat detection accuracy reaches 95% (p<0. It has been demonstrated that the traffic prediction error is reduced by 21% compared to the ARIMA benchmark (MAE=8.2Mbps) [3], and that the response delay is reduced to less than 5 seconds (a decrease of 55%) [4].It is noteworthy that during the 10Gbps DDoS attack stress test, the system's detection accuracy decreased by 18 percentage points, thus underscoring the urgent need for an enhanced joint learning mechanism [5]. The present study proffers both theoretical support and practical examples for the development of an adaptive network security system capable of effectively addressing novel intelligent attack threats.