This paper introduces an ensemble-based network anomaly detection system that synergizes classical machine learning classifiers with dimensionality reduction to balance detection accuracy and computational efficiency.  The proposed system integrates preprocessing, feature engineering, hybrid learning, and ensemble decision-making to achieve robust anomaly detection and attack categorization. We precisely evaluate five algorithms—K-Nearest Neighbor (KNN), Naïve Bayes (NB), Random Forest (RF), AdaBoost, and Gradient Boosting (GBC)—both as standalone models and within a soft-voting ensemble framework. To address high-dimensionality challenges in cybersecurity data, we adapted Principal Component Analysis (PCA) to retain 95% feature variance while reducing dimensionality by 54% (41 → 19 features), achieving a 38% latency improvement without compromising critical attack detection. A dual-phase SMOTE strategy mitigates class imbalance, enabling 100% recall for rare U2R attacks. Extensive experiments on the KDD CUP99 benchmark demonstrate the ensemble’s superiority: 93.7% accuracy (vs. 77.7–90% for individual models) and 22ms/inference latency. In addition, while Gradient Boosting achieved the highest individual average performance at 90%, the proposed ensemble exhibited strong performance in adversarial testing, gaining 97.1% accuracy compared to Gradient Boosting’s 85.2% against GAN-generated attacks. These findings establish a foundation for adaptive cybersecurity systems that employ machine learning to tackle emerging adversarial defense mechanisms, highlighting accuracy and operational feasibility in evolving threat landscapes.