A Comprehensive Analysis of Attack Parameters and Vulnerability Mapping for Enhanced Threat Detection in Network Logs Using Deep Learning

Abstract

Modern cyber-attacks exploit network vulnerabilities through dynamic patterns that often evade traditional detection models. While prior work has emphasized detection accuracy, limited research explores the relationship between extracted log parameters and real-world vulnerabilities. This study addresses that gap by proposing a deep learning-based framework that analyzes how structured log features—such as packet size, anomaly score, and port activity—contribute to identifying both known and unknown threats. The framework utilizes the NetLogFusion dataset, integrating diverse logs from routers, firewalls, and host systems. Through a multi-stage pipeline comprising log preprocessing, LaukiLogParser-based parsing, feature extraction and ranking, and detection using the Adaptive  Dual-Attention  Temporal  Convolutional  Network (ADATCN) model, the system delivers accurate and interpretable threat detection. Key features are ranked for their importance in classification, and high-impact patterns are mapped to CVEs (e.g., CVE-2013-5211, CVE-2021-44228) and MITRE ATT&CK techniques to enrich threat context. ADATCN achieves 92% AUC with a low false positive rate of 5%, demonstrating strong performance in both detection and contextual attribution. This work highlights the value of parameter-aware anomaly detection in enhancing model explainability, operational relevance, and proactive vulnerability awareness, with future directions focused on integrating live threat intelligence for real-time defense