Robust Privacy-Preserving Federated Learning for 6G Network Resource Allocation

Download

Volume 7 Issue 1, 2026

Author(s):

Nizamuddin Maitlo* Institute of Computer Science, Shah Abdul Latif University, Khairpur, nizamuddin.cs@gmail.com

Samina Rajper Institute of Computer Science, Shah Abdul Latif University, Khairpur, samina.rajper@salu.edu.pk

Manahil Shaikh Institute of Computer Science, Shah Abdul Latif University, Khairpur, manahilshaikh36@gmail.com

Hina Kareem Institute of Computer Science, Shah Abdul Latif University, Khairpuri, hinakareem.mahar@gmail.com

Abstract Sixth-generation (6G) radio access networks must allocate resources across heterogeneous service slices—enhanced Mobile Broadband (eMBB), Ultra-Reliable Low-Latency Communication (URLLC), massive Machine-Type Communication (mMTC), and Extended Reality (XR)—while meeting strict service-level targets and preserving privacy. We present a robust, privacy-compatible federated learning (FL) pipeline that learns slice-aware admission and serving decisions from gNB-side telemetry without centralizing raw logs. Using eight non-IID clients with 150 telemetry windows per client sampled every 90 s, we train a compact client-side classifier for 20 global rounds under partial participation (approximately 85%). Labels are defined by next-window SLA satisfaction (downlink throughput ≥ 40 Mbps and p95 latency ≤ 50 ms). The global model improves from 0.63 to 0.91 accuracy and from 0.62 to 0.85 macro-F1, while the training loss drops from 0.92 to 0.18. Fairness, measured as the gap between the best and worst mean per-client accuracy, is 0.08, with most per-round gaps falling in the 0.05–0.14 range. A heterogeneity snapshot confirms strong non-IID conditions across PRB utilization, CQI, p95 latency, UE density, and slice-mix entropy. The pipeline outputs reproducible CSV logs and figures and is designed to remain compatible with client-level differential privacy (norm clipping with Gaussian noise), secure aggregation, and over-the-air (AirComp) update fusion to reduce uplink time. These results indicate that FL can achieve accurate and fairness-aware 6G resource allocation under realistic participation and stringent SLAs while supporting privacy and communication optimizations.
Keywords 6G, federated learning, radio access networks, network slicing, resource allocation, non-IID data, fairness, differential privacy, secure aggregation, over-the-air computation.
Year 2026
Volume 7
Issue 1
Type Research paper, manuscript, article
Recognized by Higher Education Commission of Pakistan, HEC
Category
Journal Name ILMA Journal of Technology & Software Management
Publisher Name ILMA University
Jel Classification --
DOI -
ISSN no (E, Electronic) 2790-590X
ISSN no (P, Print) 2709-2240
Country Pakistan
City Karachi
Institution Type University
Journal Type Open Access
Manuscript Processing Blind Peer Reviewed
Format PDF
Paper Link https://ijtsm.ilmauniversity.edu.pk/arc/Vol7/i1/pdf1.pdf
Page 1-8