Vertical Federated Learning Sample Alignment

Description

VFL sample alignment is a fundamental stage in the Vertical Federated Learning (VFL) process, introduced in 3GPP Rel-20 and specified in TS 23.482 and TS 23.700. In VFL, multiple participants (VFL clients) each hold different features (columns) for the same set of samples (rows). For the collaborative training to be meaningful, all participants must agree on which samples are being used and ensure that their local datasets are aligned row-wise. Sample alignment is the procedure that achieves this, enabling consistent training and inference across participants on the same sample space.

The need for sample alignment arises from the distributed nature of VFL. Each VFL client may have data about different entities (e.g., UEs, sessions, or subscribers) and may use different identifiers or timestamps. Without alignment, the server would combine features from different samples, leading to incorrect model updates. The alignment process typically involves privacy-preserving record linkage (PPRL) techniques, such as private set intersection (PSI), to identify the intersection of samples across clients without revealing the actual identifiers. Once the common sample set is determined, each client filters its local dataset to include only those samples.

Architecturally, sample alignment is orchestrated by the VFL server, which coordinates the clients and may provide a common sample identifier or a secure protocol for intersection. The process can be synchronous or asynchronous, depending on the VFL framework. In 3GPP, the alignment stage is part of the overall VFL workflow, which includes initialization, sample alignment, training, and inference. The specifications define the interactions between VFL clients and the VFL server during this stage, ensuring interoperability.

Technically, sample alignment must address several challenges: scalability (aligning millions of samples efficiently), privacy (not leaking which samples are held by which client), and heterogeneity (different clients may have different data formats or missing samples). The 3GPP specifications provide a framework for these procedures but may reference external standards for the cryptographic protocols. The alignment stage is crucial because errors here propagate through the entire training process, degrading model accuracy.

In Rel-20, the definition of VFL sample alignment is: 'A stage in a VFL process during which samples are aligned among VFL participants, ensuring that training and inference are performed consistently across participants on the same sample space.' This definition highlights the goal of consistency and the scope of the stage. As VFL becomes more prevalent in 5G and beyond for use cases like network optimization, fraud detection, and personalized services, sample alignment will remain a key enabler for privacy-preserving collaborative AI.

Purpose & Motivation

VFL sample alignment exists to solve the problem of data alignment in distributed machine learning where different parties hold different features of the same entities. In many real-world scenarios, the same user or device is represented in multiple datasets with different identifiers (e.g., a UE may have an IMSI in one system and a temporary ID in another). Without a way to align these records, VFL cannot proceed because the model would be trained on mismatched samples.

Traditional approaches to data integration required sharing raw identifiers, which poses privacy risks. Sample alignment techniques, particularly those based on private set intersection, allow parties to find common samples without revealing their entire datasets. This is essential for compliance with privacy regulations and for maintaining business confidentiality.

The introduction of VFL sample alignment in 3GPP Rel-20 reflects the maturation of VFL in the standards. Earlier releases (Rel-19) defined the VFL client and server but did not explicitly address the alignment stage. As VFL use cases expanded, the need for a standardized approach to sample alignment became clear. The Rel-20 definition provides a common understanding for vendors and operators implementing VFL, ensuring that multi-vendor VFL deployments can interoperate. It also sets the stage for further enhancements in future releases, such as more efficient alignment protocols and support for dynamic sample sets.