A Model Drift Detection and Adaptation Framework for 5G Core Networks

TL;DR

This paper tackles model drift in 5G core networks by integrating a drift detection module with an online drift adaptation component within a functional NWDAF-enabled 5G core prototype. The authors implement an LSTM-based regressor to predict next packet length and develop a drift framework where drift is signaled when the observed squared error exceeds a threshold defined as $\mu + n \sigma$, with $n=2$ in experiments. The main contributions are the emulation of concept-drift events on real network hardware, a drift-detection mechanism that characterizes drifted concepts, and an online adaptation module with persistent and non-persistent memory modes that retrain the model to restore pre-drift performance. Results show a substantial reduction in drift alarms and demonstrate learning across unseen concepts, highlighting the framework's practical potential for maintaining reliable intelligent automation in dynamic 5G networks. Overall, the work advances robust, data-driven management and orchestration in 5G+ environments by enabling responsive model maintenance in the presence of evolving user behavior.

Abstract

The advent of Fifth Generation (5G) and beyond 5G networks (5G+) has revolutionized the way network operators consider the management and orchestration of their networks. With an increased focus on intelligence and automation through core network functions such as the NWDAF, service providers are tasked with integrating machine learning models and artificial intelligence systems into their existing network operation practices. Due to the dynamic nature of next-generation networks and their supported use cases and applications, model drift is a serious concern, which can deteriorate the performance of intelligent models deployed throughout the network. The work presented in this paper introduces a model drift detection and adaptation module for 5G core networks. Using a functional prototype of a 5G core network, a drift in user behaviour is emulated, and the proposed framework is deployed and tested. The results of this work demonstrate the ability of the drift detection module to accurately characterize a drifted concept as well as the ability of the drift adaptation module to begin the necessary remediation efforts to restore system performance.

Figures (7)

• Figure 1: Experiment System Model
• Figure 2: Drift Detection and Adaptation Framework Process Map
• Figure 3: UE-to-Server Interaction Concepts
• Figure 4: UE-to-Server Interaction Inter-Arrival Time
• Figure 5: Drift Alarm Concept 2