Abstract: With the growing adoption of deep learning in network traffic classification,model vulnerability to minor perturbations has become a critical security concern.Adversarial example attacks significantly compromise the reliability of real-world deployments.To address the longstanding trade-off between accuracy and robustness in conventional adversarial training,this paper proposes a collaborative adversarial defense framework that integrates ensemble learning with a dynamic weight constraint mechanism.The proposed approach assigns higher training weights to samples near decision boundaries to improve the model's sensitivity to vulnerable instances,while leveraging multi-model ensemble strategies to mitigate gradient overfitting and enhance robustness.Comprehensive experiments conducted on three widely used network traffic datasets－USTC-TFC2016,NSL-KDD,and CIC-IDS2017－demonstrate that the proposed method consistently improves robustness by over 10% across various attack types,and even exceeds 20% under high-intensity perturbations,compared with conventional baselines.The proposed framework de-monstrates exemplary scalability,affording seamless adaptation to heterogeneous network architectures and deployment contexts,thereby manifesting substantial practical utility and promising engineering applicability.

Key words: Adversarial examples, Adversarial training, Weight constraint, Ensemble learning, Network traffic classification