FL Explainer

FAQ

What is federated learning?
Federated Learning (FL) is a machine learning paradigm where multiple clients — devices or organizations — collaboratively train a shared global model without sending their raw data to a central server. Each client trains a local copy on its own data and transmits only the updated model parameters back to an aggregator, which combines them into a new global version. Raw data never leaves the device or institution that owns it.
How does federated learning work step by step?
A standard FL round has six steps.
Step 0: clients agree on the global model architecture and the server initializes weights.
Step 1: the server distributes the current global model to all clients.
Step 2: each client trains the model locally on its own data.
Step 3: clients return updated parameters to the server.
Step 4: the server aggregates updates (typically via FedAvg) into a new global model.
Step 5: steps 1–4 repeat until convergence.
Why is federated learning needed?
FL solves three problems at once: privacy regulations (GDPR, CCPA, PIPEDA, LGPD, 152-FZ) restrict cross-organizational data transfer; data may be too large or fragmented to centralize practically; and central data stores create high-value targets for attackers. By training where the data lives and exchanging only model parameters, FL enables ML on data that could not legally or practically be collected centrally.
What kinds of organizations use federated learning?
Federated learning is used wherever sensitive data must remain in place. Real deployments: banks and insurers training joint credit scoring without sharing customer records; hospitals collaborating on diagnostic models without moving patient data (DICOM, EHR); smartphone fleets learning keyboard predictions without uploading typing data; pharmaceutical companies running drug-discovery models on partner data without IP exposure; IoT device manufacturers improving on-device models without bandwidth costs.
What's the difference between federated learning and centralized ML?
In centralized ML, all training data is gathered into one server (or one data lake), and a model is trained on the consolidated dataset. In federated learning, the data stays distributed across many devices or organizations, and the model — not the data — moves between parties during training. Centralized ML is faster and simpler but requires data transfer; FL is slower and more complex but preserves data sovereignty.
Is federated learning the same as edge AI?
Related but not identical. Edge AI means running ML models on local devices (smartphones, IoT sensors) instead of in the cloud — typically for inference. Federated learning is a way to train models across distributed clients. Most edge AI uses centrally-trained models deployed to devices; FL is one way to train edge AI models without centralizing the training data. The two combine in scenarios like smartphones learning typing predictions across millions of devices.
What is the Guardora FL Explainer tool?
The FL Explainer is an interactive walkthrough that visualizes the Federated Learning process step by step — clients training locally, the server aggregating updates, and the global model evolving without raw data transfer. It's designed for ML engineers, product managers, and decision-makers learning federated learning concepts. The tool is part of Guardora's broader privacy-preserving ML platform, which includes Guardora VFL (vertical federated learning) and supporting infrastructure for production deployments.
How does Guardora help with federated learning?
Guardora builds commercial federated learning infrastructure: Guardora VFL for two-party vertical federated learning (used in credit scoring, fraud detection, healthcare); supporting cryptographic protocols including Paillier homomorphic encryption (1024-bit) for gradient protection; encrypted gRPC for inter-party communication; and tested integrations with XGBoost and other tabular ML models. Validated workloads include credit scoring (ROC AUC ≈ 71.3 on GBDT, 300K records under 9 minutes) and joint bank+insurance scoring (0.817 → 0.975 accuracy with 92.66% reduction in critical errors).