Practical AI-Based Traffic Classification for Next-Generation Service-Aware RAN

Modern mobile networks handle diverse traffic such as video streaming, cloud gaming, and XR/VR, making real-time analysis tailored to each service's characteristics essential. Service-aware RAN can optimize radio resource allocation and improve UE (terminal) power efficiency by identifying traffic patterns and adaptively controlling RRC (Radio Resource Control) states.

However, existing AI-based traffic classification methods require high computational cost and large-scale labeled data, making them difficult to apply in CU (Central Unit) environments with strict real-time and resource constraints. Moreover, continuously maintaining labeled datasets in response to the emergence of new services also poses a practical limitation.

To address this, we propose a practical framework that combines clustering-based pseudo-labeling with lightweight flow-level inference. This approach can effectively infer service characteristics through statistical and temporal features even in encrypted traffic.

The proposed technique is applied to RRC state control, reducing terminal power consumption while minimizing signaling overhead. Its effectiveness has been demonstrated through testbed experiments, and it can be extended to various service-aware functions in future intelligent 6G networks.