Vertical Federated Learning Client

Description

A VFL client (Vertical Federated Learning client) is a functional entity in the 3GPP Vertical Federated Learning framework, specified in TS 23.482 and TS 23.700. In VFL, multiple parties (clients) each hold different features (columns) of the same set of samples (rows). The VFL client is responsible for performing local training and inference on its own data partition, sending the results (e.g., gradients, embeddings, or model updates) to the VFL server, and updating its local parameters based on information received from the server. This collaborative approach allows multiple organizations to train a shared model without exposing their raw data, addressing privacy and regulatory constraints.

Architecturally, a VFL client can be an NWDAF (Network Data Analytics Function) or an Application Function (AF). When the client is an NWDAF, it leverages the 5G core network's analytics framework to access network data and perform federated learning tasks. When it is an AF, it typically resides outside the core network, interacting with the 5G System via the NEF. The VFL server orchestrates the training process, coordinating multiple clients, aggregating their updates, and distributing the global model or intermediate representations. The communication between clients and server is secured and may use APIs defined in 3GPP specifications.

The operation of a VFL client involves several stages. First, sample alignment ensures that all clients are working on the same set of samples (e.g., same UEs or sessions). Then, during the forward pass, each client computes intermediate results (e.g., embeddings) from its local features and sends them to the server or to other clients, depending on the VFL protocol (e.g., split learning, vertical federated learning with secure aggregation). The server combines these intermediate results to compute the loss and gradients, which are then sent back to the clients for local parameter updates. This process iterates until the model converges.

Key technical challenges addressed by the VFL client include data privacy, communication efficiency, and heterogeneity. Privacy is preserved because raw data never leaves the client; only model updates or intermediate representations are exchanged. Communication efficiency is critical because VFL can involve large models and frequent updates; techniques like gradient compression and asynchronous updates are employed. Heterogeneity arises when clients have different data distributions or computational capabilities; the VFL framework must accommodate these differences.

In Rel-19, the definition of VFL client was: 'An NWDAF or AF that holds the local dataset and performs local training and inference as asked by VFL Server. There can be multiple VFL Clients in VFL training and inference.' In Rel-20, the wording was revised to: 'A VFL participant that performs local training and inference, sends results to the VFL server, and updates local parameters based on information received from the VFL server.' This change emphasizes the client's active role in sending results and updating parameters, and generalizes the entity from specifically NWDAF/AF to any VFL participant, reflecting a broader applicability.

Purpose & Motivation

The VFL client exists to enable collaborative machine learning across multiple organizations or network functions without requiring them to share their raw data. In many scenarios, data is siloed: different operators, enterprises, or network functions hold complementary data about the same users or events, but privacy regulations (e.g., GDPR) and business confidentiality prevent data sharing. Traditional centralized machine learning would require pooling all data into a single location, which is often infeasible or illegal.

Vertical Federated Learning solves this by allowing each participant to keep its data local and only share model updates or intermediate representations. The VFL client is the entity that performs the local computation, making this privacy-preserving collaboration possible. It was introduced in 3GPP Rel-19 as part of the NWDAF and AF enhancements for federated learning, addressing the growing demand for AI/ML in network automation and optimization.

Before VFL, approaches like horizontal federated learning (HFL) assumed that all participants had the same features but different samples. However, in many telecom scenarios (e.g., predicting user experience from data held by different network functions or operators), the features are split across entities while the samples (users) overlap. VFL fills this gap. The VFL client's evolution in Rel-20, with revised wording, reflects the maturation of the concept and its broader applicability beyond just NWDAF/AF to any VFL participant.