Federated Learning Elevates Mobile Network Intelligence: Architecture & Demo

Introduction

The rapid proliferation of IoT devices, sensors, and smartphones generates massive heterogeneous data, creating a need for distributed, privacy‑preserving AI in mobile communication networks. Centralized machine learning is limited by privacy regulations, data silos, and bandwidth constraints.

Federated Learning (FL) Overview

FL enables collaborative model training without exposing raw data. Participants keep local data and exchange only encrypted model updates. The process consists of two phases: encrypted model training (aggregation of encrypted gradients) and model inference (using the globally aggregated model). FL addresses GDPR, China’s Cybersecurity Law, Data Security Law, and Personal Information Protection Law.

Relevant Standards

IEEE P3652.1 (FL architecture and application guide, 2018‑2021)

3GPP R17 TR23.700‑91 (FL use cases for 5G, 2020)

ITU‑T projects on FL for IoT and smart cities (2021)

China AIOSS reference FL architecture (2019) and CCSA research projects (2021)

FL Network‑Intelligence Architectures

Two canonical FL topologies are considered:

Client‑Server (coordinated) architecture : a central aggregator distributes the global model, collects encrypted updates from participants, aggregates them, and redistributes the improved model.

Peer‑to‑peer (decentralized) architecture : participants communicate directly, exchanging encrypted intermediate results without a central coordinator.

Both can be combined in a three‑layer hierarchy—domain FL, cross‑domain FL, and global FL—to support collaboration among users, operators, and industry stakeholders.

Customer Experience Management (CEM) Case Study

Scenario

The case evaluates Emotional Connection Scoring (ECS) and CEM across ten regions in Chongqing. Data sources include OSS (network faults) and BSS (business events). Each region acts as an FL node (NWDAF) in a horizontal FL setting.

Deployment Architecture

One region is selected as the initial coordinator; the remaining regions become distributed nodes. Each node collects local data from UPF and OMC, trains a local model, encrypts gradients and loss values, and sends them to the coordinator.

FL Workflow

1. Region A initiates an FL request and becomes the coordinator.
2. Coordinator sends initial model parameters and training constraints to Regions B, C, ….
3. Each participant trains a local model on its data, encrypts gradients and loss, and returns them to the coordinator.
4. Coordinator aggregates encrypted updates (e.g., using secure aggregation), updates the global model, and redistributes it.
5. Steps 3‑4 repeat until convergence.
6. After training, the global model is deployed locally for inference.

Algorithmic Details

Let x_b, y_b be the feature‑label pairs at node B and x_c, y_c at node C. Participants share a common neural‑network predictor F(x,θ) and use mean‑square loss L = (F(x,θ) - y)^2. Model updates are protected with a homomorphic encryption scheme [[·]] so that only aggregated results are decrypted by the coordinator.

Algorithm 1 (Federated Parameter Training):

Initialize global model parameters θ⁰.
For each communication round t = 1, 2, …
    Coordinator broadcasts θᵗ⁻¹ to all participants.
    Each participant i:
        - Compute local gradient g_i = ∇_θ L(F(x_i,θᵗ⁻¹), y_i).
        - Encrypt g_i → [[g_i]].
        - Send [[g_i]] to coordinator.
    Coordinator aggregates encrypted gradients:
        [[ḡ]] = Σ_i [[g_i]] (homomorphic addition).
        Decrypt ḡ.
        Update global model: θᵗ = θᵗ⁻¹ - η·ḡ.
    End For
Output final θ.

Experimental Setup

Data split: 90 % for training, 10 % for testing per region.

Baseline: local training without FL.

FL variants: IID (independent and identically distributed) vs. non‑IID data partitions.

Results

FL vs. non‑FL : Across all ten regions, FL‑trained models achieve significantly higher prediction accuracy than locally trained models.

IID vs. non‑IID : Models trained on IID data outperform those trained on non‑IID data, highlighting the sensitivity of FL to data heterogeneity.

These results confirm that FL can integrate disparate data sources, improve model performance, and preserve user privacy.

Conclusion and Future Work

The study demonstrates the feasibility of applying federated learning to mobile communication networks through both centralized and decentralized architectures. Future research should address non‑IID data challenges, develop more efficient secure aggregation protocols, and contribute to standardization efforts to broaden FL adoption in telecom environments.