While pre-trained large models have achieved state-of-the-art performance in network traffic analysis, their prohibitive computational costs hinder deployment in real-time, throughput-sensitive network defense environments. This work bridges the gap between advanced representation learning and practical network protection by introducing Traffic-MoE, a sparse foundation model optimized for high-efficiency real-time inference. By dynamically routing traffic tokens to a small subset of specialized experts, Traffic-MoE effectively decouples model capacity from computational overhead. Extensive evaluations across three security-oriented tasks demonstrate that Traffic-MoE achieves up to a 12.38% improvement in detection performance compared to leading dense competitors. Crucially, it delivers a 91.62% increase in throughput, reduces inference latency by 47.81%, and cuts peak GPU memory consumption by 38.72%. Beyond efficiency, Traffic-MoE exhibits superior robustness against adversarial traffic shaping and maintains high detection efficacy in few-shot scenarios, establishing a new paradigm for scalable and resilient network traffic analysis.

翻译：尽管预训练大模型在网络流量分析中已取得最先进的性能，但其高昂的计算成本阻碍了其在实时、高吞吐量敏感的网络防御环境中的部署。本研究通过引入Traffic-MoE，弥合了先进表征学习与实际网络防护之间的鸿沟。Traffic-MoE是一种专为高效实时推理优化的稀疏基础模型。通过将流量令牌动态路由至一小部分专用专家模块，Traffic-MoE有效地将模型容量与计算开销解耦。在三个面向安全的任务上进行的大量评估表明，与领先的稠密模型相比，Traffic-MoE在检测性能上实现了高达12.38%的提升。至关重要的是，它带来了91.62%的吞吐量提升，将推理延迟降低了47.81%，并将峰值GPU内存消耗削减了38.72%。除了效率优势，Traffic-MoE在面对对抗性流量整形时表现出卓越的鲁棒性，并在少样本场景下保持了高检测效能，从而为可扩展且具有弹性的网络流量分析确立了新范式。