Contrastive Adversarial Agents for Intentional Drift Induction in Network Intrusion Detection Systems

Abstract

Retraining and drift-adaptation methods for Network Intrusion Detection Systems (NIDS) are effective when adversarial behavior evolves gradually or remains predictable. However, these defenses often degrade when faced with coordinated or intentionally diverse attackers that induce abrupt and adversarial concept drift; a particularly critical vulnerability in client-server NIDS deployments for high-stakes domains such as healthcare. In this work, we investigate a stronger threat model in which an ensemble of reinforcement learning attackers collaborates to strategically disrupt adaptive defenses. We introduce a contrastive learning framework that trains multiple adversarial agents to learn distinct and non-redundant packet perturbation strategies. By encouraging behavioral diversity while maintaining attack effectiveness, the ensemble generates heterogeneous evasion patterns that are deployed sequentially or adaptively to destabilize retraining mechanisms. This approach enables systematic study of intentional, multi-modal adversarial drift and exposes vulnerabilities in standard NIDS adaptation pipelines. Our results demonstrate that contrastively trained attacker ensembles significantly reduce detection accuracy, highlighting the need for more robust and diversity-aware defensive strategies.