FL Explainer
Product
Guardora VFL

Guardora VFL

boosts machine learning model performance

Guardora VFL supports training of multiple model architectures:

Gradient Boosted Decision Trees
Logistic Regression

Guardora VFL leverages federated learning technology to enhance ML model accuracy by training on previously inaccessible external data — without the data ever leaving its owner's secure perimeter.

By implementing Guardora VFL, businesses increase revenue through more powerful predictive models.

The solution specializes in tabular data processing.

Explore the Guardora VFL API documentation for implementation details.

Easy to install and start using

Guardora VFL API

A fragment of Swagger UI with Guardora VFL API:

node

get

/node

Get Node Info

post

/node

Post Node

delete

/node

Delete Node

get

/node/public_key

Generate Node Public Key

post

/node/start

Start

post

/node/stop

Stop

remotes

get

/remotes

List Remotes

post

/remotes

Add Remote

delete

/remotes/{remote_name}

Remove Remote

get

/remotes/{remote_name}/check_connection

Check Connection

projects

post

/projects/{project_name}

Create

Delivery format

Access to the repository

How to install

On each participating party perform the following steps:

  • Clone the repository
  • Configure the environment
  • Build and run Docker containers

Intuitive REST API

Use the product’s friendly API to perform model training and inference. Or integrate it into your ML-pipeline

Who plays?

You test solo or with partners

Data

Your data, as well as data from your partners. We can provide our test data

Infrastructure and Security

On-premise or private cloud.

The client itself establishes the necessary security level, and obtains full control over data and processes

Scalability

Scalability within the client's internal infrastructure

Pricing

“Pay-As-You-Infer” licensing model with postpayment - monthly licensing fee based on the number of useful requests (received predictions from federatively trained model)

Support

Guardora’s support is included in the license fee

Typical Logical Architecture of the Solution

Dataset synchronization (Private Set Intersection, PSI), training, validation, inference (with homomorphic encryption)

Dataset synchronization (Private Set Intersection, PSI), training, validation, inference (with homomorphic encryption)

Method of obtaining the data table (e.g., CSV file, SQL connection to a database). Implemented by the client

Homomorphic encryption

Method of obtaining the data table (e.g., CSV file, SQL connection to a database). Implemented by the client

Required for high-load inference >300 RPS

Required for high-load inference >300 RPS

Dataset synchronization (Private Set Intersection, PSI), training, validation, inference (with homomorphic encryption)

Dataset synchronization (Private Set Intersection, PSI), training, validation, inference (with homomorphic encryption)

Method of obtaining the data table (e.g., CSV file, SQL connection to a database). Implemented by the client

Method of obtaining the data table (e.g., CSV file, SQL connection to a database). Implemented by the client

Homomorphic encryption

Required for high-load inference >300 RPS

Required for high-load inference >300 RPS

Main use cases of Guardora VFL

Training scoring and anti-fraud
models on other parties’ data

Tasks to be solved:

  • Increasing scoring accuracy using partners’ data without direct access to it
  • Detecting complex fraudulent schemes on cross-corporate data

For whom:

Credit bureaus, Scoring providers, Banks and microfinance organizations, Insurance companies, Payment systems, Telecom operators

Creating joint ML products

Tasks to be solved:

Developing new products that require combining data from several companies

For whom:

Scoring providers, Fintech companies

Using data within corporate groups

Tasks to be solved:

Overcoming internal barriers to data exchange between companies within a group
Creating unified models for the entire group

For whom:

Fintech holdings, Groups of companies, including cross-border

Conducting pilot projects and retro-tests

Tasks to be solved:

Rapid testing of hypotheses on real data of potential clients
Demonstration of the value of the product without the risk of data leakage

For whom:

Scoring providers, Providers of solutions for working with personal data

Guardora VFL Licensing

Guardora VFL allows training ML models on data from multiple companies while preserving the confidentiality of each company's data. Using a credit scoring model as an example, this means that:

  • 01/ The model receives more diverse data, which contributes to better generalization ability
  • 02/ Each company maintains control over its data
  • 03/ Ultimately, the increase in model accuracy (for example, by 1% in GINI) translates to more accurate risk assessment

With a better model the client company receives additional revenue through more accurate customer segmentation, reduced defaults, and optimized credit decisions.

Based on the described effect, Guardora offers a transactional “Pay-As-You-Infer” licensing model with postpayment with the following parameters:

  • 01/ Monthly licensing fee based on the number of useful requests (received predictions from federatively trained model)
  • 02/ The more requests, the lower the cost per request
  • 03/ Technical support cost is included in the licensing fee

An example for various business segments

FAQ

What is Guardora VFL?
Guardora VFL is a commercial Vertical Federated Learning platform built for two-party scenarios in regulated industries. It lets two organizations — typically a bank/insurer with target labels and an analytics or data vendor with complementary features — jointly train a tabular ML model without transferring raw data or labels between perimeters. Each party runs Guardora VFL inside its own infrastructure; the platform handles secure ID alignment (Private Set Intersection), federated training, optional gradient encryption (Paillier 1024-bit), and inference serving.
What ML models does Guardora VFL support?
Guardora VFL is purpose-built for tabular (structured) data. Currently supported model architectures are Gradient Boosted Decision Trees (GBDT) — including XGBoost-compatible workflows — and Logistic Regression. These cover the majority of production ML use cases in banking, insurance, telecom, and fintech (credit scoring, fraud detection, churn prediction, customer segmentation, risk underwriting). GBDT is the workhorse — tested on 78K + 97K record bank+insurance benchmark (0.817 → 0.975 accuracy) and on 300K credit-scoring records (ROC AUC ≈ 71.3, training <9 minutes).
How is Guardora VFL deployed?
Guardora VFL is delivered as a Docker-based on-premise installation that each participating party runs inside its own perimeter. The standard install procedure is: clone the repository, configure environment variables, build and run the Docker containers. Communication between parties happens over encrypted gRPC. The platform is also exposed via REST API — explore the live API documentation at https://apidoc.guardora.ru. Each party retains full control over data and infrastructure security level.
What's the pricing model — "Pay-As-You-Infer"?
Guardora VFL uses a Pay-As-You-Infer transactional licensing model with postpayment. You pay a monthly fee based on the number of useful inference requests (predictions returned from the federatively-trained model) — not on training time or seat count. The more requests, the lower the cost per request. Technical support is included in the licensing fee. There is no upfront license cost; pilots and demos are available before commercial commitment. Final pricing is provided after a discovery call — contact form on the page.
Who uses Guardora VFL — typical customer profiles?
Four customer archetypes.
Credit bureaus, scoring providers, banks, microfinance organizations, insurance companies, payment systems, telecom operators — for scoring and anti-fraud models on partner data.
Scoring providers and fintech companies — for creating joint ML products by combining feature sets from multiple companies.
Fintech holdings and corporate groups (including cross-border) — for unblocking ML training across subsidiaries that are legally separate.
Scoring providers and personal-data SaaS vendors — for running retro-tests on prospective client data during pilots without leakage risk.
What's the throughput / performance of Guardora VFL?
Inference performance scales with deployment. Standard inference reaches well above 100 RPS on commodity hardware; GPU-equipped deployments exceed 300 RPS (the page architecture diagram explicitly calls out this threshold for high-load scenarios). On the credit-scoring benchmark, multi-threaded inference reached 650 requests per second — approximately 0.008 seconds per request, comparable to non-federated XGBoost. Training time depends on encryption mode: GBDT on 300,000 records takes under 9 minutes without encryption, ~1.4 hours with Paillier homomorphic encryption on 50,000 records.
Is technical support included?
Yes. Technical support is included in the monthly licensing fee for the Pay-As-You-Infer model — there is no separate support contract. Guardora's engineering team assists with deployment, integration with existing ML pipelines, performance tuning, and incident response. For on-premise deployments, the level of operational support is agreed in the commercial contract.
How do I try Guardora VFL?
Three paths.
Demo Guardora VFL — a guided demo of the product (https://guardora.ai/product/demo-guardora-vfl/).
Demo Veils — a separate demo of Guardora's functional-transformation tool.
Contact form on the product page for a discovery call to discuss your specific use case, run a pilot on your actual data, and receive a commercial proposal.

Write To Us

To test the demo version of the product and find out the pricing conditions, fill out the form and we will contact you