Federated Learning Server

Description

The FL Server (Federated Learning Server) is a functional entity defined in 3GPP specifications 28.105 and 28.858 that orchestrates the federated learning process within the management domain. Its primary responsibility is to aggregate the ML model updates received from multiple FL Clients into a single global ML model. The FL Server is the coordination point of a federated learning system: it selects participants, distributes the current global model, collects local updates, and combines them to improve the model iteratively. Crucially, the FL Server never receives the raw training data from the clients — only their model updates — which is what makes the overall process privacy-preserving.

Architecturally, the FL Server resides in the management plane and is typically associated with a Management Data Analytics (MDA) or AI/ML management function. It exposes management services that allow FL Clients to register, to be selected for training rounds, and to submit their updates. The FL Server maintains the state of the global model, tracks the progress of training rounds, and applies aggregation algorithms. The most common aggregation method is federated averaging (FedAvg), in which the global model is updated as a weighted average of the client updates, with weights often proportional to the amount of local training data. More advanced aggregation schemes may be used to handle heterogeneous clients, stragglers, or malicious participants.

The FL Server drives the federated learning workflow in rounds. At the start of a round, it selects a subset of eligible FL Clients based on criteria such as availability, data quality, or resource constraints. It then distributes the current global model to the selected clients. Each client trains locally and returns a model update. Once a sufficient number of updates are collected — or a timeout expires — the FL Server aggregates them to form a new global model. This model is then used in the next round. The process repeats until convergence or a configured stopping condition. The FL Server is also responsible for evaluating the global model, potentially using a validation dataset or by requesting evaluation results from clients.

From a management and security standpoint, the FL Server must handle several challenges. It must ensure that client updates are authentic and have not been tampered with, which may involve authentication and integrity protection. It may apply secure aggregation so that it cannot inspect individual updates, further strengthening privacy. It must also manage fairness and robustness, for example by detecting and mitigating poisoning attacks from malicious clients. 3GPP management specifications define the relevant capabilities for configuring and monitoring the FL Server, including the ability to set aggregation parameters, round schedules, and client selection policies.

The FL Server is a cornerstone of 3GPP's AI/ML management architecture. By aggregating distributed learning without centralizing data, it enables operators to build accurate models from data spread across many network elements while respecting privacy and regulatory constraints. It is the counterpart to the FL Client and together they form the standardized federated learning enabler introduced in Release 19 and refined in Release 20.

Purpose & Motivation

The FL Server exists to coordinate and aggregate the contributions of many distributed FL Clients into a single, high-quality global model. Without a coordinating server, federated learning would be impossible: there would be no entity to distribute the model, collect updates, and combine them. The FL Server solves the problem of how to learn from data that is scattered across numerous network elements and domains without ever centralizing that data. It provides the aggregation logic that turns many local model updates into a globally useful model.

Historically, 3GPP's AI/ML management work began with centralized approaches, where training data is collected at a central function. While simple, this approach struggles when data is privacy-sensitive, too large to transfer, or distributed across administrative domains. Federated learning was introduced to overcome these limitations, and the FL Server is the central enabler of that paradigm. It allows the network to benefit from distributed data while keeping that data local, aligning with privacy regulations and reducing bandwidth consumption.

The FL Server also provides the standardization anchor for multi-vendor interoperability. By defining a standard server function, 3GPP ensures that FL Clients from different vendors can participate in a common federated learning process orchestrated by a single FL Server. This is essential for operators who deploy multi-vendor networks and want to leverage AI/ML across their entire infrastructure. Introduced in Release 19 and refined in Release 20, the FL Server is a key building block for autonomous, AI-driven network management.