Optimization of multi-modal appearance detection network security situation awareness model

Abstract

Addressing the issues of low accuracy, high false positive rates, and poor generalizability in traditional cybersecurity situational awareness models regarding heterogeneous data fusion, complex threat identification, and dynamic situation assessment, this paper proposes a multimodal appearance-based detection cybersecurity situational awareness model (MM-ADN).This model fuses three types of heterogeneous data—network traffic, system logs, and topology images—to capture security features across different dimensions via a multimodal feature extraction module. It employs a cross-modal attention mechanism to achieve deep feature fusion and combines this with a temporal appearance detection network to perform situation assessment and prediction. Experiments conducted on the UNSW-NB15, CIC-IDS2018 datasets, achieving accuracy rates of 97.8% and 96.5%, respectively, with F1 scores of 0.976 and 0.952, and false positive rates as low as 1.20% and 0.77%. The model significantly outperforms traditional single-modal and mainstream multi-modal models in terms of detection accuracy, dynamic adaptability, and robustness against interference, and can provide technical support for proactive cybersecurity defense.