Optimized federated learning framework for Open Radio Access Networks (O-RAN)

Abstract

The Open Radio Access Network (O-RAN) architecture represents a transformative paradigm for realizing the grand vision of beyond fifth-generation (B5G) and sixth-generation (6G) mobile networks. Through its foundational characteristics of disaggregation, standardized open interfaces enabling genuine multi-vendor interoperability, and embedded intelligence via RAN Intelligent Controllers (RICs), O-RAN enables unprecedented network capabilities. The programmable nature of O-RAN, particularly through the Non-Real-Time RIC (Non-RT-RIC) for strategic optimization and Near-Real-Time RIC (Near-RT-RIC) for tactical resource management, creates a platform where artificial intelligence and machine learning (ML) can revolutionize network operations. Through these advancements, O-RAN enables the vision of mobile networks simultaneously supporting heterogeneous service requirements such as enhanced Mobile Broadband (eMBB), ultra-Reliable Low-Latency Communications (uRLLC), and massive Machine-Type Communications (mMTC). Realizing O-RAN’s embedded intelligence based use-cases critically depends on the ML model training framework itself. Federated Learning (FL), a distributed machine learning paradigm enabling collaborative model training across disaggregated network nodes without centralizing raw data, emerges as a promising solution. Unlike centralized learning, FL is uniquely suited for O-RAN’s multi-vendor, privacy-sensitive environment where: operational data cannot be shared across vendor boundaries due to confidentiality and competitive concerns; massive data volumes from millions of user equipment and thousands of base stations create prohibitive communication bottlenecks if transmitted to central servers; and regulatory frameworks like GDPR mandate data localization. FL enables critical O-RAN use cases including distributed anomaly detection across multi-vendor equipment for security, collaborative spectrum optimization without exposing proprietary algorithms, privacy-preserving quality-of-experience prediction using sensitive subscriber data, and real-time radio resource management leveraging localized channel state information that would be stale if centrally processed. However, deploying FL in O-RAN environments presents formidable technical challenges fundamentally distinct from traditional federated learning scenarios. O-RAN’s strict control loop timing requirements makes learning time equally critical as model accuracy. Severe resource constraints exist at multiple layers: Distributed Unit (O-DU) and Radio Unit (O-RU) have limited computational capacity as edge nodes; Near-RT-RICs operate on resource-constrained servers; and backhaul links connecting these components have time-varying, bandwidth-limited connectivity. Moreover, the mobile devices participating in FL training are frequently handed over between O-DUs, causing lost computations, unpredictable changes in the participant set, and aggregation delays from straggler nodes. Production O-RAN deployments also demand concurrent execution of multiple FL tasks where each task potentially serving different network slices (eMBB, uRLLC, mMTC) with conflicting objectives and quality-of-service requirements while competing for shared infrastructure resources for the model training. Together, it poses three main challenges. The trained FL model must be (i) efficient to handle diverse system capabilities including heterogeneous compute power and varying channel conditions; (ii) robust to ensure global model performance under optimized aggregation algorithms; and (iii) reliable to converge within the defined thresholds. This thesis systematically addresses these challenges through three progressive contributions. First, we develop a unified resource-efficient and communication-efficient FL framework that jointly solves trainer selection and resource allocation while attacking the communication bottleneck through synergistic integration of momentum-based acceleration and aggressive compression techniques. Second, we propose MHORANFed, a mobility-aware Hierarchical Federated Learning framework that explicitly handles inter-O-DU handover disruptions through hierarchical aggregation mapping to O-RAN’s architecture, adaptive resource reallocation for mobile participants, and dynamic aggregation weight adjustment. Third, we present the O-FL rApp, a system-level orchestration framework operating in the Non-RT-RIC that manages the complete lifecycle of multiple concurrent Federated Multi-Agent Reinforcement Learning tasks through strategic task and slice assignment, tactical resource reallocation among competing tasks. Unlike prior works focusing on individual aspects in isolation or idealized settings, our integrated solution demonstrates practical, production-grade deployment feasibility through rigorous theoretical analysis, extensive experimental validation using benchmark FL datasets, realistic O-RAN testbeds with 3GPP-compliant channel models and 5G traffic traces, aligned with O-RAN Alliance specifications. This research establishes foundational principles and operational frameworks enabling autonomous, adaptive, privacy-preserving federated learning framework at scale which are essential building blocks for realizing the full potential of intelligent O-RAN in B5G evolution and 6G networks.

Date	19 Dec 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Kim Khoa Nguyen (Supervisor)