Glossary

Feature Owner

A feature owner is a participant in a Vertical Federated Learning (VFL) system that possesses a subset of the features (input variables) for an aligned set of entities but does not hold the target labels.

Get in touch Learn more

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

VERTICAL FEDERATED LEARNING

What is a Feature Owner?

A core participant in a privacy-preserving, decentralized machine learning system.

A feature owner is a participant in a Vertical Federated Learning (VFL) system that possesses a distinct, non-overlapping subset of the features (input variables or columns) for an aligned set of entities (e.g., customers, devices) but does not hold the labels (target values). This entity contributes its local feature data to a collaborative model training process without ever exposing the raw data to other parties. Its role is defined by its partial view of the complete feature space and its reliance on secure protocols to compute on its data slice.

During training, a feature owner typically hosts a segment of a split neural network, performing vertical forward propagation on its local features to produce an intermediate output at the cut layer, which is then securely transmitted. It also receives gradients via vertical backpropagation to update its model portion. The feature owner's operational constraints and incentives are central to system design, influencing communication overhead, secure aggregation protocols, and the overall vertical training protocol.

VERTICAL FEDERATED LEARNING

Core Responsibilities of a Feature Owner

In a Vertical Federated Learning (VFL) system, a Feature Owner is a critical participant who holds a specific subset of input features for a shared set of entities. Their responsibilities are defined by the need to contribute to model training while rigorously protecting their proprietary data.

Local Feature Computation

The Feature Owner executes the forward pass on their segment of the split neural network using their local features. This involves computing the model's layers up to the pre-defined cut layer. The resulting intermediate output (or embedding) is the only data transmitted from their system, never the raw input features. This computation must be efficient and compatible with the overall vertical training protocol.

Secure Gradient Handling

During vertical backpropagation, the Feature Owner receives gradient signals from the downstream party (typically the Label Owner). They must use these gradients to compute updates for their local model segment via vertical gradient computation. These updates are often protected using vertical secure aggregation or homomorphic encryption before being shared to prevent inference of other parties' data or the labels from the gradient values.

Data Alignment & Integrity

Before training begins, the Feature Owner must participate in entity alignment to establish the common set of samples across parties. This is often done using a Private Set Intersection (PSI) protocol. They are responsible for maintaining the integrity and order of their local vertical data partition throughout training, ensuring their feature vectors correctly correspond to the aligned entity IDs.

Protocol Compliance & Coordination

The Feature Owner must strictly adhere to the agreed-upon vertical training protocol and vertical inference protocol. This includes:

Synchronizing computation rounds with other parties.
Adhering to communication schedules to minimize vertical communication overhead.
Implementing the correct cryptographic primitives for privacy-preserving vertical FL.
Interfacing with the chosen Vertical FL framework for orchestration.

Privacy & Security Enforcement

A primary duty is to enforce data privacy. This involves applying technical safeguards like differential privacy in federated learning by adding noise to intermediate outputs or gradients. They must also guard against security threats, ensuring their system is not vulnerable to attacks that could compromise their model segment or infer other parties' data, a key aspect of federated learning attack mitigation.

Resource & Performance Management

The Feature Owner manages the vertical computation overhead and local infrastructure costs. This includes:

Optimizing the local forward/backward pass for their hardware.
Managing the latency and bandwidth costs associated with transmitting encrypted intermediates.
Ensuring system availability to participate in training rounds, contributing to overall system reliability and convergence speed.

VERTICAL FEDERATED LEARNING

How a Feature Owner Operates in VFL

In Vertical Federated Learning (VFL), a feature owner is a critical participant that holds a specific subset of the input data. This entry details its operational role, responsibilities, and interactions within the secure training protocol.

A feature owner is a participant in a Vertical Federated Learning (VFL) system that possesses a distinct, non-overlapping subset of features (input variables) for the aligned set of entities but does not hold the training labels. Its primary operational function is to locally compute the forward pass of its segment of a split neural network on its private features, generating an intermediate output or embedding. This output is then securely transmitted—often in encrypted form—to the label owner or a coordinating server to continue the forward propagation and loss calculation for the global model.

During backpropagation, the feature owner receives gradient signals from the label owner corresponding to its part of the model. It uses these gradients to compute updates for its local model segment via vertical gradient computation. The feature owner must adhere to the defined vertical training protocol, which governs communication rounds and enforces privacy through techniques like homomorphic encryption or secure multi-party computation (MPC). This ensures raw feature data is never exposed, maintaining data sovereignty while contributing to the collaborative model.

VERTICAL FEDERATED LEARNING

Real-World Examples of Feature Owners

A Feature Owner is a critical participant in a Vertical Federated Learning (VFL) system. These examples illustrate the distinct roles and data assets held by different organizations in collaborative, privacy-preserving model training.

Retail Bank (Credit Features)

A bank acts as a feature owner for a joint credit scoring model with an e-commerce platform. The bank holds financial features for shared customers, such as:

Account balances and transaction history
Loan repayment records
Credit card utilization ratios

These features are highly predictive but regulated. Using VFL, the bank computes intermediate outputs from its local model segment on this data. These encrypted results are sent to the label owner (e-commerce platform holding default labels) for aggregation, enabling accurate risk assessment without exposing raw financial data.

Hospital (Clinical Features)

In a healthcare consortium training a disease prediction model, a hospital is a feature owner for clinical data. It possesses structured electronic health record (EHR) features for a patient cohort, including:

Lab test results (e.g., HbA1c, cholesterol)
Vital signs and medication history
Diagnostic codes (ICD-10)

A separate research institute holds genomic data (another feature set), and an insurance provider acts as the label owner with disease progression outcomes. The hospital uses a split neural network to process its local features, sharing only encrypted intermediate embeddings to build a comprehensive model while maintaining HIPAA/GDPR compliance.

Telecom Provider (Behavioral Features)

A telecommunications company participates as a feature owner in a customer churn prediction project with a streaming service. The telecom holds behavioral features derived from network usage for aligned users:

Data consumption patterns and call duration
Device type and network connectivity logs
Geographic mobility patterns

The streaming service (label owner) holds subscription cancellation data. Through secure entity resolution using Private Set Intersection (PSI), the parties align their customer IDs. The telecom's local model processes its features, and only the computed gradients are securely aggregated, preventing leakage of sensitive mobility or usage data.

Automotive Manufacturer (Sensor Features)

An automaker is a feature owner in a collaborative model for predictive maintenance with a parts supplier network. The automaker collects real-time sensor data from vehicles:

Engine temperature and RPM telemetry
Brake pad wear sensor readings
Vibration and acoustic emission data

Supplier companies hold data on component material batches and manufacturing conditions (other feature sets). A central fleet operator holds the failure labels. The automaker's edge devices in vehicles perform local forward propagation on sensor data. The resulting features are used within the VFL protocol to predict part failures, optimizing supply chains without transmitting raw sensor streams.

IoT Platform (Environmental Features)

An industrial IoT platform managing smart buildings acts as a feature owner for an energy optimization model. It controls sensors that capture environmental features across multiple facilities:

Temperature, humidity, and occupancy sensor data
HVAC system runtime and setpoints
Lighting usage patterns

A utility company holds electricity pricing and grid load data, and the building management company holds the labels (total energy cost). The IoT platform's gateway devices compute on the vertically partitioned sensor data. Using homomorphic encryption, the encrypted outputs are sent for aggregation, enabling dynamic pricing response without revealing proprietary operational data.

Payment Processor (Transactional Features)

A payment processor serves as a feature owner in a federated fraud detection system with merchant partners. It possesses granular transactional features:

Transaction amount, time, and frequency
Merchant category codes (MCC)
Device fingerprinting and IP geolocation data

E-commerce merchants hold cart contents and user session data (additional features), while the acquiring bank holds the chargeback labels. The payment processor applies its local model segment to transactional features. Through vertical secure aggregation protocols, its update is combined with others to detect fraud patterns that no single party could identify alone, all while keeping transaction details private.

KEY ROLES IN VERTICAL FEDERATED LEARNING

Feature Owner vs. Label Owner

A comparison of the two primary participant roles in a Vertical Federated Learning (VFL) system, highlighting their distinct responsibilities, data holdings, and computational tasks.

Role & Responsibility	Feature Owner	Label Owner
Primary Data Held	A subset of features (input variables) for the aligned entities.	The target labels (output values) for the aligned entities.
Core Objective in VFL	To contribute feature-specific knowledge to improve the global model while keeping raw feature data private.	To orchestrate training, compute the primary loss, and often own the final prediction task.
Model Architecture Segment	Holds and trains the bottom layers of a split neural network, up to the cut layer.	Holds and trains the top layers of a split neural network, from the cut layer to the output.
Key Computation per Round	Performs a forward pass on local features to compute an intermediate output; computes gradients for its local model segment during backpropagation.	Completes the forward pass using received intermediate outputs to compute loss; initiates backpropagation and coordinates gradient aggregation.
Data Privacy Risk	Risk of feature data leakage via intermediate outputs or gradients. Mitigated by encryption or perturbation.	Risk of label information leakage. May also learn sensitive patterns from aggregated intermediate inputs.
Communication Overhead	Transmits encrypted intermediate outputs (and potentially gradients) each round. Bandwidth scales with cut layer size and batch size.	Receives and aggregates intermediate outputs from all feature owners; broadcasts loss gradients. Often the communication hub.
Inference Role	Computes and securely transmits the intermediate output for its feature set for each prediction query.	Receives intermediate outputs, completes the forward pass, and returns the final prediction to the querying party.
System Incentive	Gains utility from an improved model without centralizing sensitive feature data (e.g., a bank's credit features).	Gains a powerful predictive model leveraging cross-party features without owning them (e.g., an e-commerce platform's purchase labels).

VERTICAL FEDERATED LEARNING

Frequently Asked Questions

A feature owner is a core participant in a Vertical Federated Learning (VFL) system. These questions address their role, responsibilities, and the technical mechanisms that enable secure collaboration.

In practical terms, imagine a bank and an e-commerce company collaborating on a credit risk model. The bank holds financial features (e.g., account balance, loan history), while the e-commerce company holds behavioral features (e.g., purchase history, browsing patterns). Both have data on the same customers (aligned entities), but with different features. Here, each organization is a feature owner. Their core function is to compute on their local features as part of a distributed split neural network, generating intermediate outputs or gradients that are securely shared to facilitate joint model training without exposing their raw, proprietary data tables.

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

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.

Talk to Us

Feature Owner

What is a Feature Owner?