1. Federated Learning and FedOps at a Glance

📍 Where you are: Basics 1/3 — ① Federated learning and FedOps → ② Data and model → ③ Setup

🎯 What you do in this step: Get the big picture before touching anything. Once you know why federated learning is needed, who does what, and which part each FedOps screen handles, you will see “why I am clicking this button now” in the later steps.

  1. The problem federated learning solves
  2. Two roles: Task Owner and Client
  3. FedOps components
  4. The life of a Task in FedOps
  5. Next

The problem federated learning solves

A good model needs a lot of varied data. But sensitive data, such as medical or financial records, cannot be gathered in one place. Laws, regulations, and contracts forbid it, and even when gathering is possible the leak risk is hard to accept. Across countries it is even harder. So if each hospital builds a model only from its own data, the data is small and skewed toward that hospital’s patients, and performance suffers.

Federated Learning does the opposite. Instead of gathering the data, it sends the model to the data.

flowchart LR
  S["Aggregation server<br/>Global Model"] -->|"① Send current model"| C1["Hungarian hospital<br/>local data"]
  S -->|"①"| C2["Cleveland hospital<br/>local data"]
  C1 -->|"③ Trained parameters only"| S
  C2 -->|"③"| S
  1. The server sends the current model to each participating institution (client).
  2. Each client trains that model a bit further on its own data, inside its own PC.
  3. Only the trained parameters (the numbers inside the model) go back to the server. The data never leaves the PC.
  4. The server merges the parameters into one (aggregation) to make a new Global Model.

One such loop is a round. Repeating rounds makes the Global Model behave as if it had seen every institution’s data (ideally).

📚 Federated learning background — what is hard
  • Data differs from institution to institution (non-IID). The two hospitals in this tutorial are an example. 38% of the Hungarian hospital’s patients have heart disease, versus 46% at Cleveland, and age and test value distributions differ too. Each client pulls the model toward its own data, so a simple merge can give a model that fits one side poorly. This tutorial does not create this difference on purpose; you experience it as a difference that real hospital data already has.
  • Communication is expensive. The whole model travels every round. The number of rounds and of participating clients is the cost.
  • Participants are not always there. A client’s PC can be off or lose the network. The server waits until the required number of clients has gathered.
  • Sending only parameters is not perfect privacy. Attacks that infer parts of the training data from parameters have been studied. Techniques such as differential privacy and secure aggregation exist separately; this tutorial does not cover them.

Two roles: Task Owner and Client

 Task OwnerClient (participant)
DoesDecides what to learn, builds and publishes the code and first model, runs the trainingJoins a published Task and trains locally on its own data
DecidesData format, model structure, number of rounds, number of participants, who to acceptWhether to join, what data to put in
Data heldData for design and the first model (may have none). In this tutorial, the Hungarian hospital’s recordsIts own institution’s data. In this tutorial, the Cleveland hospital’s records
Main screensFedOps Console + Agent StudioAgent Studio (+ Registry)
ReceivesThe Global Model and overall metricsIts own participation record and metrics

One account can play both roles. When an Owner also joins its own Task as a Client, the screen shows an Owner · Participant badge. This is how you practice alone.

FedOps components

NameWhereWhat it doesWho
FedOps WebWeb (home page)Introduction and docs. Open Console at the top right takes you to the ConsoleEveryone
FedOps ConsoleWebCreate Tasks, publish to the Registry, approve participants, run the aggregation server and campaigns, monitoringMostly the Owner
Agent StudioMy PC (Docker)Edit code, local training, readiness checks, join federated learning, build Agents and APIsEveryone
Federated TaskConsole + StudioThe unit of federated learning. A code bundle containing the model, preprocessing, training code, and Tool definitionCreated by the Owner
RegistryConsole / StudioThe list of published Tasks. Clients find Tasks here and request to joinEveryone
Aggregation server (FL server)FedOps cloudRuns the rounds and merges parametersStarted by the Owner
Global ModelConsoleModel versions produced by federated learning (v1, v2, …)Visible to everyone

Why Studio runs on my PC: the data must not leave my PC. Training happens in Studio (my PC); only the merging happens on the server.

📚 Where does aggregation happen? — FedOps servers

Aggregation — receiving the parameters each participant sends and merging them — happens on a FedOps server. FedOps is not a service tied to one place: the same platform can be installed and run on different infrastructure.

FedOps serverWhere aggregation runs
Main FedOpsThe on-premises Kubernetes cluster of the CCL lab at Gachon University — the one this tutorial uses
AWSAmazon Web Services cloud
Kakao CloudKakao Cloud

Whichever server you use, data stays on each participant’s PC and only trained parameters go to the server. An institution can pick where aggregation runs to fit its data policy or budget. You choose which server to connect to in the FedOps server field on the Agent Studio sign-in screen (3. Setup).

The life of a Task in FedOps

This is the storyline the whole tutorial follows. Note who changes each state.

flowchart LR
  A["Web Draft<br/>(Owner, Console)"] --> B["Workspace<br/>code · data<br/>(Owner, Studio)"]
  B --> C["Release Candidate<br/>(Owner, Studio)"]
  C --> D["Published Release<br/>+ Initial Model<br/>(Owner, Console)"]
  D --> E["Join request · approval<br/>(Client ↔ Owner)"]
  E --> F["Campaign run<br/>(Owner server + Client training)"]
  F --> G["Global Model vN"]
  G --> H["Agent Tool AI<br/>(model use)"]
  G -->|"next campaign"| F

Next

2. The Practice Data and the AI Model — what per-hospital data and model this tutorial trains, and why they were chosen.


Back to top

© 2026 Cognitive Computing Lab, Gachon University