Vertical Federated Learning Server

Description

The VFL server is a core architectural component introduced in 3GPP Release 19 to enable Vertical Federated Learning (VFL) within the 5G system. VFL is a machine learning paradigm where multiple parties hold different features (or different parts of a model) for the same set of samples, and collaborate to train a joint model without sharing raw data. The VFL server acts as the central coordinator for this process. It can be implemented as a Network Data Analytics Function (NWDAF) or an Application Function (AF), depending on the deployment scenario and the nature of the participating clients. Its primary responsibilities span both the training and inference phases of the VFL lifecycle.

During the VFL training process, the VFL server discovers and selects suitable VFL clients, which are typically UEs, network functions, or application servers that hold local data or model components. Once clients are selected, the server orchestrates the training rounds. In each round, clients perform local computations on their private data and send intermediate results—such as gradient information or loss values—to the server. The server integrates these local training results, computes aggregated gradient or loss information, and sends the updates back to the clients. This iterative process allows the global model to improve without exposing raw data, thereby preserving privacy and reducing communication overhead. The server ensures that only one VFL server exists per VFL process, maintaining a single point of coordination and consistency.

In the VFL inference process, the VFL server aggregates local inference results from the VFL clients. Each client computes a partial inference based on its local model component and data, and the server combines these partial results to generate the final VFL inference result. This final result is then delivered to the consumer, which could be an application or a network function requesting the inference. The server thus acts as both a coordinator and an aggregator, ensuring that the distributed inference is correctly assembled.

The VFL server interacts with other 5G core network functions through standardized interfaces. For example, when implemented as an NWDAF, it may use the Nnwdaf services to communicate with other NFs, and it may leverage the NEF for external AF interactions. The server also relies on discovery mechanisms to identify potential VFL clients, which may involve querying the NRF or using direct provisioning. Security and privacy are paramount; the server must ensure that only authorized clients participate and that data exchanged during training and inference is protected.

Overall, the VFL server is a key enabler for privacy-preserving, distributed machine learning in 5G and beyond. It allows operators and service providers to leverage data across multiple domains without centralizing sensitive information, unlocking new use cases in areas such as network optimization, personalized services, and vertical industry applications.

Purpose & Motivation

The VFL server exists to enable Vertical Federated Learning in 3GPP networks, addressing the growing need for collaborative machine learning across multiple data owners without compromising data privacy or requiring massive data centralization. Traditional centralized machine learning approaches require pooling data from various sources into a single location, which raises significant privacy, regulatory, and bandwidth concerns. VFL overcomes these by allowing parties to keep their data local while still training a joint model. However, VFL introduces coordination complexity: clients must synchronize their computations, exchange intermediate results, and aggregate them correctly. The VFL server solves this by providing a single, standardized coordination point.

Before the introduction of the VFL server in Release 19, 3GPP had defined other federated learning architectures, such as horizontal federated learning (HFL) where clients share the same feature space but different samples. VFL is more complex because clients hold different features for the same samples, requiring careful alignment and secure aggregation. The VFL server was created to manage this complexity, offering functions for client discovery, selection, training orchestration, and inference aggregation. It also ensures that only one server exists per VFL process, avoiding conflicts and ensuring consistency.

The VFL server is motivated by use cases in network automation, where data is distributed across RAN, core, and edge domains. For instance, optimizing network slicing may require combining RAN performance data with core network load data, which are owned by different entities. The VFL server allows such cross-domain learning without moving raw data, thus respecting ownership boundaries. It also supports regulatory compliance by keeping data local. As 5G evolves towards 6G, the VFL server is expected to become a fundamental building block for distributed AI/ML in telecom networks.