Adversarial attacks pose significant threats to intrusion detection models in cloud data centers. Previous defense strategies, including adversarial training, adversarial example detection and mitigation, and model ensemble voting, exhibit limited robustness, often failing to generalize. These approaches degrade detection performance on normal traffic and struggle to adapt to dynamic network changes due to their reliance on model-specific or attack-specific designs and static ensembles. To address these limitations, we propose a robustness optimization mechanism for intrusion detection models based on dynamic ensemble learning. We construct a heterogeneous pool of adversarially trained models, leveraging diverse algorithms to reduce adversarial attack transferability. A confidence-based dynamic model selection algorithm then selects optimal model combinations from the pool in real-time, adapting to evolving network traffic. Finally, an ensemble decision-making strategy aggregates predictions from the selected models using a voting mechanism, enhancing detection accuracy and ensuring robust defense against dynamic adversarial threats. Experimental results show that our approach outperforms single models and static ensembles, maintaining strong accuracy on normal traffic while demonstrating superior resilience in dynamic adversarial scenarios.