AI-UPF: A Tool for Mining Traffic Value

Since early 2024, 5G has entered the 5G-Advanced (5G-A) era. As user traffic growth slows and the value derived from basic connectivity reach its limit, networks are shifting from rapid expansion to a phase of high-quality and stable development. Consequently, operators are transitioning from traffic-based to experience-based operation.

Experience-Based Operation Solution

ZTE provides an experience-based operation solution powered by "AI+" connectivity intelligence. It provides customized experience guarantee based on user levels, services, and scenarios, unlocking the commercial value of 5G-A. By introducing AI, the user plane function (UPF) enables deep service identification and accurate experience measurement, providing a basis for differentiated service operation at the upper layer. The network data analytics function (NWDAF) conducts in-depth analysis of user services and service measurement data, and generates user assurance policies based on the real-time load of both wireless and core networks, ensuring end-to-end user experience. Fig. 1 shows the solution.

This article focuses on the introduction of AI-UPF in the 5G-A core network to build key intelligent capabilities, enabling precise application identification and accurate service experience perception, providing a data foundation for differentiated services, and ultimately supporting personalized, end-to-end experience assurance.

AI-UPF Unlocks Traffic Value

The AI-UPF is an NE that incorporates AI technology to optimize and enhance the traditional UPFs. With a built-in AI engine and model invocation support, it delivers two key capabilities: intelligent service identification and intelligent experience measurement, unraveling the data pipeline to fully extract network traffic value and support experience-based operation.

Intelligent Service Identification with Greater Scope and Accuracy

As 5G-A networks and generative AI rapidly advance, applications are constantly evolving and new ones are emerging at a fast pace. Traditional signature-based identification requires rule updates with each application change, making timely identification of new or updated applications difficult. In addition, with growing data security concerns and the widespread use of new encryption technologies, traditional methods are increasingly inadequate for experience-based operation. AI-UPF intelligent service identification effectively solves these challenges.

The AI-UPF automates service analysis through its built-in AI engine, shortening  the signature library update cycle from monthly to daily. In traditional service identification, once key features such as domain names are extracted from unknown traffic, manual analysis is required to determine the correlation between the features and their corresponding applications, slowing the update cycle. The AI-UPF invokes a large model via its built-in AI engine and uses the model’s semantic understanding capabilities together with local knowledge graph entries to analyze correlations—such as similarity and keywords—among key features of service flows. It then infers the applications behind unknown traffic and cluster the traffic accordingly. Based on the common characteristics within each cluster, the model obtains the application name, type, and description, enabling intelligent tagging and near real-time updates of the signature library.

The AI-UPF uses AI to subdivide services and accurately identify encrypted traffic or private protocols. When sub-services within an application are encrypted, plaintext features are unavailable. Different types of service flows appear as packet sequences with different spatiotemporal features (e.g., packet length, number of packets,

and timestamps), making simple rule-based identification methods ineffective. By building an AI model that learns the spatiotemporal characteristics of encrypted traffic, the AI-UPF can identify sub-services (e.g., video, audio, and live broadcast) within applications. To ensure model generalization and identification accuracy, model training and fine-tuning are critical. Training samples are generated through an automatic sampling system that monitors version changes of mainstream applications on the network, triggers service dialing tests and labeling, and accumulates tens of millions of samples. With continuous training and optimization, the model achieves an accuracy of 95% or above in identifying encrypted services, supporting personalized and specific services assurance.

Intelligent Experience Measurement, Closer to Real User Experience

Traditional service quality evaluation relies on network transport-layer KPIs (such as packet loss, jitter, and delay). With the evolution of application transmission technologies—such as encryption and dynamic bitrate—these KPIs no longer effectively reflect the service quality at the application layer.

To accurately predict user-perceived quality, the AI-UPF builds multiple measurement models, such as jamming detection, delay detection, and bitrate detection, to extract data that closely reflect real user experience from weak indicators such as packet length, time intervals, and uplink/downlink characteristics in encrypted applications. For model training, more than 20 degraded network quality scenarios are simulated in the lab using network impairment instruments. Real terminals are used to automatically test more than 100 mainstream applications, generating massive sample data for targeted training of various measurement models, thereby ensuring measurement accuracy. Once deployed in live networks, these models achieve over 90% accuracy in user experience measurement. When degradation occurs, the issue is promptly reported to the NWDAF, triggering end-to-end service experience guarantee mechanisms.

To meet the high performance and low latency requirements of intelligent service processing, the AI-UPF introduces GPU hardware to speed up model inference. It achieves millisecond-level inference latency through data parallelism and multi-GPU parallel processing, increasing performance by tens of times compared to general-purpose CPUs. It also provides service-based model inference with unified inference interfaces, simplifying the inference process and enabling the deployment of more intelligent applications in the future.

Built on the AI-UPF and network intelligence plane, ZTE's 5G-A core network experience-based operation solution enables real-time collection and analysis of user service quality and refined service assurance, boosting operator revenue during the stable development phase of 5G-A networks.