GEN-DRIFT: Generative AI-driven drift handling for beyond 5G networks

Abstract

Beyond fifth-generation (B5G) networks enable high data rates, low latency, and massive machine communications, driving digital transformation across sectors. The integration of Artificial Intelligence and Machine Learning (AI/ML) technologies plays a vital role in enhancing the performance and efficiency of B5G networks. However, the dynamic and ever-evolving service demands associated with B5G use cases lead to the occurrence of drift, which can significantly degrade the performance of AI/ML models. Drift occurrence often results in violations of Service Level Agreements (SLAs) and over- or under-provisioning of resources, ultimately impacting user experience and network reliability.

Drift detection and adaptation are essential for addressing the dynamic service demands of B5G networks. Existing threshold approach and various other frameworks, have significant limitations, — SLA violations from delayed drift detection and inefficient resource management due to frequent retraining. This paper proposes a drift handling framework that determines drift promptly after its occurrence using Generative Artificial Intelligence (Gen-AI). The proposed Gen-AI framework is evaluated for a Quality of Service Prediction use case on the Open Radio Access Network (O-RAN) Software Community (OSC) platform and compared to the existing threshold and other frameworks. Also, a real-time dataset from the Colosseum testbed is considered to evaluate the Network Slicing (NS) use case with the proposed Gen-AI framework for drift handling.

The results demonstrate that the proposed Gen-AI framework leverages both Generative Adversarial Network (GAN) and Variational AutoEncoder (VAE), significantly enhances drift detection and adaptation time in B5G networks. Specifically, in the QoS prediction use case, GAN achieves 98% drift detection accuracy, while the VAE achieves 95% , compared to 85% for the classifier framework, 25% for the threshold-based approach. In addition, a similar kind of results is observed in case of the network slicing use case. These results highlight the effectiveness of the proposed Gen-AI framework in proactively handling drift with reduced detection and adaptation time, making it a promising solution for B5G networks.