MEGATRON: Machine Learning in 5G with Analysis of Traffic in Open Radio Access Networks

With the advent of SG and next-generation cellular networks and the increasing complexity of assigning users traffic types for efficient resource allocation, Open Radio Access Networks (O-RAN) offer intelligent virtualized frameworks for optimizing network operations related to supporting diverse types of traffic. In this paper, we utilize the native support for machine learning in O-RAN to develop a transformer-based SG traffic classification system that identifies, with high accuracy, conditions when broadband, machine-to-machine type communication, and ultra-reliable low-latency communication are present. By utilizing distinct temporal slices of O-RAN-defined key performance indicators generated from traffic captures as inputs (as opposed to directly accessing user-plane data) and filtering for non-critical control traffic, we ensure user confidentiality while maintaining a high degree of classification performance. Our transformer model is able to achieve an average offline accuracy of 99%+ for the longest traffic slice length, with the online deployment achieving an average of similar to 90% accuracy across all slice lengths.