Federated Learning Client

Description

The FL Client (Federated Learning Client) is a functional entity defined in 3GPP specifications 28.105 and 28.858 that participates in a federated learning process within the management domain of a mobile network. Federated learning is a distributed machine learning paradigm in which multiple participants collaboratively train a shared global model without exchanging their raw local data. The FL Client is the participant-side function: it receives a global model (or model initialization) from the FL Server, performs local training on its own data, and returns only the model outcome — the model updates, gradients, or weights — to the FL Server. The raw training data never leaves the FL Client's domain, which is the fundamental privacy guarantee of the approach.

Architecturally, the FL Client sits within the management plane, typically as part of a Management Data Analytics (MDA) or AI/ML management framework. It interacts with the FL Server over a service-based management interface. The FL Client may be instantiated in various network domains — for example, in a RAN node, a core network function, or a network slice management entity — depending on where the local data resides and where the model is intended to be applied. The FL Client is responsible for the local training loop: it applies the received global model to its local dataset, runs the configured optimization algorithm (e.g., stochastic gradient descent) for a number of local epochs, and produces the model update. It may also perform local validation and quality checks before submitting the update.

The operation of an FL Client follows a round-based protocol. In each federated learning round, the FL Server selects a subset of available FL Clients and distributes the current global model. Each selected FL Client trains locally and returns its update. The FL Server then aggregates these updates — commonly via federated averaging — to produce an improved global model, which is redistributed in the next round. This iterative process continues until the model converges or a stopping criterion is met. The FL Client therefore acts as both a consumer of the global model and a producer of local model updates.

From a management perspective, the FL Client must be discoverable, selectable, and monitorable by the FL Server. 3GPP management specifications define the relevant capabilities, including the ability to register as an FL participant, to report training status and metrics, and to be configured with training parameters such as batch size, learning rate, and number of local epochs. Security and privacy are paramount: model updates may be protected using secure aggregation, differential privacy, or encryption so that even the FL Server cannot inspect individual client contributions in detail.

The FL Client is a foundational building block for AI/ML-driven network automation in 3GPP. It enables operators to leverage data distributed across many network elements — data that often cannot be centralized due to privacy, regulatory, or bandwidth constraints — while still producing accurate, globally useful models. As networks move toward higher levels of autonomy, the FL Client role becomes increasingly important for use cases such as anomaly detection, traffic prediction, energy saving, and mobility optimization.

Purpose & Motivation

The FL Client exists to solve a fundamental tension in AI/ML for telecommunications: the most valuable training data is often distributed across many network elements and domains, but centralizing that data is frequently impractical or prohibited. Data may be subject to privacy regulations, may be too voluminous to transfer economically, or may be commercially sensitive to the operator or its customers. Traditional centralized machine learning requires moving all training data to a single location, which creates privacy risks, bandwidth costs, and single points of failure.

Federated learning addresses this by inverting the data flow: instead of bringing data to the model, the model is brought to the data. The FL Client is the embodiment of this inversion on the participant side. It allows each data owner to keep its raw data local while still contributing to a shared model. This preserves privacy by design, reduces data transfer requirements, and enables the use of data that would otherwise be inaccessible.

Within 3GPP, the FL Client was introduced as part of the broader effort to standardize AI/ML management enablers. Earlier releases defined centralized AI/ML management, where training data is collected at a central function. While effective, this approach has clear limitations for distributed, privacy-sensitive, or high-volume data. The FL Client, introduced in Release 19, provides the standardized participant-side function needed to make federated learning interoperable across multi-vendor networks, ensuring that clients from different vendors can participate in a common federated learning process orchestrated by an FL Server.