BALANCING ACCURACY AND ROBUSTNESS IN INTRUSION DETECTION: AN ENSEMBLE DEEP LEARNING APPROACH

Intrusion Detection Systems increasingly face adaptive, multimodal cyberattacks, yet most of the current systems are highly accurate in clean conditions and inadequately robust to adversarial perturbations, creating a major robustness gap. This challenge is necessary to ensure the current network settings where the attacks are modified at the structural, time, and content levels. The purpose of this study was to design and test a sound multimodal intrusion detection system that is able to meet accuracy and adversarial resilience. Our experimental study was based on four publicly available datasets of IDS, which combined the flow features, packet payloads, graph-structured interactions with the host telemetry into a single architecture. The model was called Robust Multimodal Graph-Transformer Ensemble (RMGTE) and contrastive alignment, adversarial training, Mixture-of-Experts routing and randomized smoothing were used as its additions to improve the robustness and was evaluated against RF, GNN and hybrid models. Results indicate that RMGTE attained 96.1% accuracy, 0.93 macro-F1, and 0.61 adversarial accuracy, which is significantly higher than baselines; furthermore, it had a certified robustness radius of 0.48, and high cross-dataset generalization (0.87 macro-F1 on BoT-IoT holdout). These results indicate that multimodal fusion and robustness-aware ensemble design makes a substantial contribution to the performance and reliability of IDS, which provides a promising future path to resilient next-generation cybersecurity systems.