SoftCoS-Net: QoS-proxy-driven fused supervision for ultra-lightweight edge CoS classification

Network Traffic Classification (NTC) is a key capability for encrypted-traffic analysis, Quality-of-Service (QoS) provisioning, and QoS-aware edge network control. However, encryption and intra-application service heterogeneity make rigid application-to-Class-of-Service (CoS) mappings prone to label noise, causing mismatches between edge-side CoS marking and downstream QoS scheduling. We present SoftCoS-Net, an ultra-lightweight edge CoS classifier trained with QoS-proxy-driven fused supervision. From single-point passive flow observations, the framework estimates QoS-proxy indicators for bidirectional flows and performs centroid-based clustering to obtain probabilistic soft labels. These soft labels are then fused with application-derived hard labels in a confidence-aware manner to form higher-fidelity supervision. SoftCoS-Net adopts a dual-branch multimodal encoder that learns byte-level protocol-syntax features and compresses flow-level statistical features, together with a sample-adaptive gated fusion module that dynamically weights the two modalities. Experiments on ISCX VPN-nonVPN2016, USTC-TFC2016, and a real-world VPN/Non-VPN Network Application Traffic (VNAT) dataset show that SoftCoS-Net achieves up to 0.998 weighted F1 and improves VNAT performance over MH-Net by 1.4% and MTC-MAE by 13.4%. Meanwhile, it uses only 0.183 M parameters, corresponding to 86.4% fewer parameters than MTC-MAE and over 99.9% fewer than MH-Net. CPU-only inference results further indicate low model-side latency and high throughput after flow-level inputs are available, suggesting the potential of SoftCoS-Net for low-overhead edge-side inference under the evaluated profiling setting.