Glossary

Vertical Federated Learning

A federated learning scenario where organizations hold data about the same entities but with different feature spaces, requiring entity alignment and split neural network training.

Get in touch Learn more

Data scientist building training data pipeline on laptop, data preprocessing visible, technical workspace.

SPLIT NEURAL NETWORK TRAINING

What is Vertical Federated Learning?

A federated learning scenario where participating organizations hold data about the same set of entities but with different feature spaces, requiring entity alignment and split neural network training.

Vertical Federated Learning (VFL) is a privacy-preserving machine learning paradigm where two or more parties hold complementary features about an overlapping set of common entities, but no single party possesses the complete feature set or the label. Unlike horizontal federated learning, which partitions data by samples, VFL partitions data by feature space, requiring secure entity alignment to match records across silos before training can begin.

The training process employs a split neural network architecture, where each participant maintains a local bottom model that transforms its private features into intermediate embeddings. These embeddings are then aggregated by a central server hosting the top model, with privacy-preserving techniques such as homomorphic encryption or differential privacy applied to the exchanged activations and gradients to prevent feature leakage between parties.

ARCHITECTURAL CHARACTERISTICS

Key Features of Vertical Federated Learning

Vertical Federated Learning (VFL) addresses the scenario where collaborating parties hold data on the same set of entities but with non-overlapping feature spaces. Unlike horizontal federated learning, VFL requires entity alignment and split neural network training to build a comprehensive model without exposing raw data.

Entity Alignment

The foundational preprocessing step where participating organizations identify overlapping samples across their silos without revealing non-matching entities. This typically employs Private Set Intersection (PSI) protocols, which allow two parties to discover the intersection of their sample IDs without disclosing the full list to each other. In genomic consortia, this often involves matching patient cohorts across hospitals that hold different modalities—one institution may possess whole-genome sequences while another holds electronic health records for the same individuals.

Split Neural Network Architecture

The defining technical mechanism of VFL where the deep learning model is partitioned across participants. Each party maintains a bottom model that processes its local features into intermediate representations called embeddings or activations. These embeddings are then exchanged and concatenated by a central top model that computes the final prediction. Critically, raw features never leave their host institution—only abstract numerical representations are shared, preserving data locality while enabling joint inference.

Privacy-Preserving Gradient Exchange

During backpropagation, the top model computes gradients that must be transmitted back to each party's bottom model. To prevent gradient leakage attacks that could reconstruct private inputs, VFL implementations often integrate:

Homomorphic Encryption: Gradients are encrypted before transmission, allowing the top model to compute on ciphertext
Differential Privacy: Calibrated noise is injected into gradients to provide mathematical privacy guarantees
Secure Multi-Party Computation: Cryptographic protocols that compute gradient updates without any party seeing another's raw values

Inference Without Data Centralization

Once trained, the VFL model operates in a distributed inference mode. When a new sample arrives, each party runs its bottom model locally on its available features, encrypts the resulting embedding, and transmits it to the aggregation server. The top model combines these embeddings to produce a prediction—for example, a disease risk score—without any single institution ever holding the complete feature vector. This architecture is essential for real-time clinical decision support systems spanning multiple hospital networks.

Asymmetric Feature Contribution

Unlike horizontal FL where all parties contribute the same feature types, VFL inherently handles heterogeneous data modalities. In a genomic VFL scenario:

Hospital A contributes: DNA sequence variants, gene expression profiles
Hospital B contributes: MRI imaging features, pathology reports
Hospital C contributes: Longitudinal lab results, medication history The split architecture naturally accommodates this asymmetry, learning cross-modal representations that would be impossible to discover if data remained isolated.

Label Ownership Asymmetry

In many VFL scenarios, only one party—typically the active party or guest—holds the ground truth labels. The other participants, called passive parties or hosts, contribute features but lack labels. This creates a unique training dynamic where the active party must orchestrate the learning process, computing losses and initiating backpropagation. In genomic applications, the label holder is often the institution with clinical outcomes or phenotype annotations, while other parties contribute molecular or imaging features.

VERTICAL FEDERATED LEARNING

Frequently Asked Questions

Clear answers to the most common technical and strategic questions about training split neural networks across institutions with different feature spaces.

Vertical Federated Learning (VFL) is a decentralized machine learning paradigm where multiple organizations hold data about the same set of entities but with different feature spaces, requiring entity alignment and split neural network training. Unlike horizontal federated learning where parties share the same features for different samples, VFL addresses scenarios where, for example, a hospital holds genomic data and a pharmacy holds prescription records for overlapping patients. The process begins with Private Set Intersection (PSI) to identify common entities without revealing non-overlapping ones. A split neural network is then trained: each party maintains a local bottom model that processes its own features into intermediate representations called embeddings or activations, which are sent to a top model—often held by a coordinator—that computes the final prediction and loss. Gradients flow back only to the respective bottom models, ensuring raw features never leave their host institution. This architecture enables collaborative model training on vertically partitioned data while preserving the confidentiality of proprietary feature sets.

FEDERATED TOPOLOGY COMPARISON

Vertical vs. Horizontal Federated Learning

Structural comparison of the two primary federated learning data partitioning scenarios for genomic consortia.

Feature	Vertical FL	Horizontal FL	Federated Transfer Learning
Data Partitioning Axis	By feature space (columns)	By sample space (rows)	By both features and samples
Sample Overlap	Same entities across parties	Different entities across parties	Minimal or no overlap
Feature Overlap	Different feature sets	Same feature space	Different feature spaces
Entity Alignment Required
Common Genomic Use Case	Multi-omics integration across biobanks	GWAS across hospital cohorts	Cross-modal knowledge transfer
Neural Architecture	Split neural network	Federated Averaging (FedAvg)	Domain adaptation networks
Privacy Mechanism	Entity resolution + encryption	Secure aggregation	Co-representation learning
Communication Pattern	Peer-to-peer intermediate activations	Client-to-server gradient updates	Hybrid parameter exchange

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

VERTICAL FEDERATED LEARNING

Real-World Applications in Genomics

Vertical federated learning enables organizations with complementary data about the same patients—such as a hospital with clinical records and a lab with genomic sequences—to collaboratively train models without exposing sensitive raw data.

Genotype-Phenotype Association

A hospital holds phenotypic data (diagnosis, lab results, imaging) while a sequencing center holds genomic data (WGS, WES) for the same patients. Vertical FL enables joint training of a model to predict disease risk from genetic variants without either party seeing the other's raw data.

Entity alignment uses hashed patient identifiers or Bloom filters for privacy-preserving record linkage
The model is split: bottom layers process each party's features locally, top layers aggregate learned representations
Enables discovery of novel genotype-phenotype correlations across institutional boundaries

2-5+

Participating Institutions

Multi-Omics Cancer Subtyping

Different research centers specialize in distinct molecular assays: one performs transcriptomics (RNA-seq), another proteomics (mass spectrometry), and a third holds clinical outcomes. Vertical FL fuses these heterogeneous feature spaces to build unified cancer subtype classifiers.

Each modality is processed by a dedicated neural network tower before encrypted intermediate representations are exchanged
The split learning architecture ensures raw omics data never leaves its host institution
Improves subtype classification accuracy by 15-30% compared to single-modality models

Drug Response Prediction

A pharmaceutical company possesses high-throughput drug sensitivity screens while a hospital network holds patient genomic profiles and treatment outcomes. Vertical FL enables joint modeling to predict which patients will respond to specific therapies.

Aligns entities across cell line experiments and patient cohorts using privacy-preserving intersection protocols
The pharma partner contributes molecular fingerprint features, the hospital contributes germline and somatic variants
Accelerates biomarker-driven clinical trial design without exposing proprietary compound data or patient records

Radiogenomic Fusion Models

A radiology department holds CT/MRI imaging data while a genomics core holds tumor sequencing data for the same cancer patients. Vertical FL trains a joint model that learns to predict genomic mutations directly from imaging features.

The imaging party runs a CNN or vision transformer locally; the genomics party runs a variant encoder
Encrypted intermediate feature vectors are exchanged via secure aggregation protocols
Enables non-invasive genomic profiling where biopsy is impractical, with mutation prediction AUCs exceeding 0.85

Federated GWAS with Covariates

Traditional federated GWAS uses horizontal FL (same features, different samples). Vertical FL extends this to scenarios where one biobank has genotype data and another has environmental exposure data or lifestyle covariates for overlapping populations.

Employs secure matrix multiplication protocols to compute cross-party covariance structures
Preserves the statistical power of joint multivariate analysis while maintaining strict data compartmentalization
Critical for studying gene-environment interactions in complex diseases like asthma and diabetes

Privacy-Preserving Polygenic Risk Scoring

A direct-to-consumer genetic testing company holds genotype arrays while an academic medical center holds longitudinal health records. Vertical FL enables joint calibration of polygenic risk scores without either party exposing their core asset.

The genetic party computes variant effect weights locally; the clinical party computes phenotype residuals
Only encrypted gradient updates flow between parties during model training
Produces risk scores with improved calibration across diverse ancestries compared to single-source models

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.