Guardians of the Network: An Ensemble Learning Framework With Adversarial Alignment for Evasive Cyber Threat Detection

Advanced cyber threats such as zero-day exploits and sophisticated evasion techniques challenge Network Intrusion Detection Systems (NIDS). To address this, we propose a robust machine learning framework that integrates multi-source data fusion, protocol-aware preprocessing, and ensemble learning. Our study uses a comprehensive dataset of 12.7 million real-world network flows (10.1M benign, 2.6M malicious) collected from enterprise environments. Our key innovation is a weighted voting ensemble—combining Logistic Regression, Decision Trees, and a 1D-CNN—which achieves 99.8% detection accuracy while reducing false positives by 4.9% compared to individual models. The system also incorporates a lightweight adversarial aligner to counter evasion techniques (e.g., IP fragmentation, MAC spoofing), recovering up to 95% of baseline recall. Notably, under extreme class imbalance (1:99), our framework maintains 80.1% recall with only 8.2 false positives per million packets, outperforming deep learning models like LSTM and 1D-CNN while using 100 times fewer parameters. These results demonstrate the framework's practicality for efficient, high-throughput NIDS deployments in real-world settings.