Ensemble-based adversarial defense with dynamic model selection for intrusion detection systems

This paper presents FEDMS (Framework for Ensemble Defense with Dynamic Model Selection), a novel adversarial defense system for network intrusion detection that addresses the vulnerability of machine learning-based security systems to adversarial attacks. Our approach integrates nine heterogeneous detection models across three categories (deep learning, traditional ML, and statistical anomaly detection) with a multi-dimensional confidence scoring mechanism that enables dynamic model selection based on input characteristics. The framework incorporates real-time adaptation capabilities and comprehensive uncertainty quantification to maintain effectiveness against evolving threats. Extensive experimental evaluation on KDD Cup 1999 and UNSW-NB15 datasets demonstrates superior performance, achieving 96.8% accuracy on clean data and maintaining 75.5% average accuracy under strong adversarial attacks, significantly outperforming existing defense mechanisms. Real-time simulation validates practical deployment feasibility with 12.4ms average processing latency. The system’s dynamic selection mechanism reduces computational overhead while maintaining robust security guarantees, making it suitable for enterprise network deployment.