Dual self-attention is what you need for model drift detection in 6G networks

Ameur, Mazene; Brik, Bouziane; Ksentini, Adlen
IEEE Transactions on Machine Learning in Communications and Networking, 4 June 2025

The advent of 6G networks heralds a transformative shift in communication technology, with Artificial Intelligence (AI) and Machine Learning (ML) forming the backbone of its architecture and operations. However, the dynamic nature of 6G environments renders these models vulnerable to performance degradation due to model drift. Existing drift detection approaches, despite advancements, often fail to address the diverse and complex types of drift encountered in telecommunications, particularly in time-series data. To bridge this gap, we propose, for the first time, a novel drift detection framework featuring a Dual Self-Attention AutoEncoder (DSA-AE) designed to handle all major manifestations of drift in 6G networks, including data, label, and concept drift. This architectural design leverages the autoencoder’s reconstruction capabilities to monitor both input features and target variables, effectively detecting data and label drift. Additionally, its dual self-attention mechanisms comprising feature and temporal attention blocks capture spatiotemporal fluctuations, addressing concept drift. Extensive evaluations across three diverse telecommunications datasets (two time-series and one non-time-series) demonstrate that our framework achieves substantial advancements over state-of-the-art methods, delivering over a 13.6% improvement in drift detection accuracy and a remarkable 94.7% reduction in detection latency. By balancing higher accuracy with lower latency, this approach offers a robust and efficient solution for model drift detection in the dynamic and complex landscape of 6G networks.


DOI
Type:
Journal
Date:
2025-06-04
Department:
Communication systems
Eurecom Ref:
8254
Copyright:
© 2025 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

PERMALINK : https://www.eurecom.fr/publication/8254