Glossary

Vertical Federated Learning

A federated learning scenario where datasets share overlapping sample spaces but differ in their feature spaces, enabling two organizations holding different attributes about the same users to collaboratively train a model without exposing their respective columns.

Get in touch Learn more

Data engineer managing feature store on laptop, feature definitions visible, casual data engineering session.

FEATURE-PARTITIONED COLLABORATION

What is Vertical Federated Learning?

A distributed machine learning paradigm where two or more entities hold different attributes about the same set of overlapping users, enabling collaborative model training without exposing raw feature columns.

Vertical Federated Learning is a privacy-preserving machine learning scenario where participating datasets share an overlapping sample space but differ in their feature spaces. Unlike horizontal federated learning, which distributes data across different users, vertical FL applies when organizations possess disjoint sets of attributes about the same individuals—such as a telecom operator holding browsing data and a bank holding transaction records for identical subscribers. The core technical challenge lies in aligning samples across silos using Private Set Intersection protocols before jointly computing a model without any party revealing its proprietary columns to the others.

The training process typically employs a split neural network architecture with a privacy-preserving entity resolution layer. Each participant maintains its own bottom model to process local features, transmitting only encrypted intermediate embeddings to a neutral third-party server that computes the forward pass and backpropagates gradients. Security is enforced through homomorphic encryption or secure multi-party computation, ensuring that even the aggregator cannot reverse-engineer individual feature values. This architecture is critical for telecom operators collaborating with financial institutions to build fraud detection models while maintaining strict data sovereignty and regulatory compliance.

ARCHITECTURAL PRIMITIVES

Key Characteristics of Vertical Federated Learning

Vertical Federated Learning (VFL) addresses the scenario where two or more institutions hold different features for the same set of entities. Unlike horizontal federated learning, VFL requires secure entity alignment and intermediate computation exchange to jointly train a model without exposing raw feature columns.

Entity Alignment via Private Set Intersection

Before training begins, participants must identify overlapping samples without revealing non-overlapping entities. Private Set Intersection (PSI) protocols enable two parties to compute the intersection of their sample IDs cryptographically, ensuring that only the common user base is used for training while keeping exclusive customers confidential.

Zero

Non-Overlapping IDs Exposed

Split Neural Network Architecture

In VFL, the model is physically partitioned across participants. Each party maintains a bottom model that processes its local features into intermediate representations called embeddings. These embeddings, not raw data, are exchanged and fed into a top model hosted by a central server or a neutral third party to compute the final prediction.

Secure Gradient and Loss Exchange

During backpropagation, the top model computes the loss and sends encrypted gradients back to each participant's bottom model. To prevent gradient leakage attacks that could reconstruct private features, VFL implementations often employ homomorphic encryption or differential privacy to protect the backward pass.

Asymmetric Feature Ownership

A defining characteristic of VFL is that one party typically holds the label data (the active party), while collaborators provide complementary features (passive parties). This asymmetry requires specialized loss computation protocols where the label owner securely shares the loss gradient without exposing the ground truth labels to passive participants.

Inference with Multi-Party Computation

Post-training inference requires real-time coordination between all parties. When a new user arrives, each participant runs its bottom model locally and encrypts the output. A Secure Multi-Party Computation (SMPC) protocol aggregates these encrypted embeddings to produce the final prediction without any single party seeing the full feature vector.

Communication-Computation Trade-off

VFL exchanges intermediate activations and gradients for every batch, creating a synchronous communication bottleneck unlike the asynchronous nature of horizontal FL. Optimizations include quantization of embeddings and reduced precision training to minimize the data transfer overhead between collaborating organizations.

TOPOLOGY COMPARISON

Vertical vs. Horizontal Federated Learning

Structural comparison of the two primary federated learning partitioning strategies based on how data is distributed across participating entities.

Feature	Vertical FL	Horizontal FL	Federated Transfer Learning
Data Partitioning Axis	By feature space (columns)	By sample space (rows)	By both feature and sample space
Overlapping Entity Space	Same users, different attributes	Different users, same attributes	Minimal overlap in users and attributes
Typical Participants	2-10 institutional silos	100s to millions of edge devices	2 distinct organizational domains
Entity Alignment Requirement	Private Set Intersection required	Not required	Limited co-occurrence alignment
Primary Cryptographic Primitive	Secure Multi-Party Computation	Secure Aggregation	Homomorphic Encryption
Communication Pattern	Peer-to-peer or coordinator-mediated	Hub-and-spoke (server-centric)	Coordinator-mediated with domain adaptation
Sample ID Overlap
Feature Space Overlap

VERTICAL FEDERATED LEARNING

Frequently Asked Questions

Clear, technically precise answers to the most common questions about training models across organizations that hold different attributes for the same set of users.

Vertical Federated Learning (VFL) is a privacy-preserving machine learning paradigm where two or more parties hold datasets with overlapping sample spaces but disjoint feature spaces, enabling collaborative model training without exposing raw data columns. Unlike Horizontal Federated Learning where participants share the same schema for different users, VFL applies when organizations possess different attributes about the same set of entities—for example, a telecom operator holding call detail records and a bank holding transaction histories for overlapping subscribers. The process begins with Private Set Intersection (PSI) to cryptographically identify common samples without revealing non-overlapping identities. Training then proceeds using split neural network architectures where each party maintains its own bottom model to process local features, exchanging only intermediate embeddings or encrypted gradients through a neutral coordinator. The core challenge lies in aligning samples across silos and computing loss gradients without any party gaining visibility into another's proprietary feature columns, typically solved through homomorphic encryption or secure multi-party computation protocols during the forward and backward passes.

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

ECOSYSTEM

Related Terms

Vertical Federated Learning relies on a specific set of cryptographic and architectural primitives to align disparate feature spaces without exposing raw data. These related concepts form the technical backbone of secure, multi-party model training.

Entity Alignment

The critical pre-processing step in Vertical Federated Learning where overlapping sample IDs are matched across datasets without revealing non-overlapping records. Techniques like Private Set Intersection (PSI) allow two telecom operators to discover common subscribers holding different attributes (e.g., billing data vs. call detail records) before training begins. Without secure alignment, feature spaces cannot be logically joined.

Pre-training

Required Stage

Split Neural Networks

A distributed architecture where the bottom layers of a deep model are partitioned across clients, each processing its own private features. Only the smashed data (intermediate activations) is exchanged, not raw inputs. In a telecom context, one base station's model might process signal strength while another handles user mobility patterns, with a central server combining these representations to predict network congestion.

Secure Multi-Party Computation (SMPC)

A cryptographic protocol enabling multiple parties to jointly compute a function over their private inputs while keeping those inputs secret. In Vertical Federated Learning, SMPC secures the aggregation of gradients and loss calculations, ensuring that no participant can reverse-engineer another's feature values from shared intermediate computations.

Feature Engineering Across Silos

The challenge of creating consistent, meaningful input variables when datasets have zero feature overlap. This involves:

Feature binning for categorical mismatches
Embedding alignment for high-cardinality identifiers
Normalization across different sensor scales Telecom operators must harmonize features like temporal granularity (hourly vs. millisecond logs) before vertical training can succeed.

Inference with Partial Features

A deployment challenge unique to Vertical Federated Learning where a trained model must generate predictions when only one party's features are available. Solutions include knowledge distillation to transfer collaborative knowledge into a unilateral student model, or maintaining a lightweight online protocol where the missing party contributes its portion of the forward pass via encrypted API call.

Label Protection

In many vertical scenarios, only one party holds the ground truth labels (e.g., a financial institution knows fraud outcomes). Techniques like label differential privacy and homomorphic encryption prevent the label-holding party from leaking supervision signals to feature-holding partners during backpropagation, preserving the competitive advantage of owning the target variable.

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.