Aggregator of model updates
AI Models & Assistants / Knowledge library
Federated Learning
Training across decentralized data without centralizing raw examples
Why learn Federated Learning?
It enables collaborative learning where data must remain distributed, while still requiring explicit privacy and security controls.
Where it came from
Federated learning coordinates local training and aggregates updates from many clients. It reduces raw-data movement but introduces heterogeneity, communication, privacy, and attack concerns.
Build your understanding
Work through these exercises in order. Try the task before opening a reference.
- Explain federated training: What stays on each client in standard federated learning?
- Describe aggregation: What does a coordinator commonly aggregate?
- Avoid a privacy misconception: Why is federated learning not automatically private?
- Handle heterogeneous clients: Why can non-IID client data make training harder?
- Evaluate fairly: Which evaluation exposes uneven federated performance?
Documentation and sources
- Primary reference (arxiv.org)
Connect the ideas
Explore next
Real World Example
Privacy-Preserving Predictive Maintenance for Industrial IoT
Manufacturing firms operating globally distributed factory equipment that requires predictive maintenance models.
The business problem
Firms cannot centralize raw sensor data due to strict data residency laws and the high bandwidth costs of transmitting massive, high-frequency vibration logs from remote factory sites to a central data center.
How this topic solves it
Federated Learning enables the training of a shared predictive maintenance model by keeping raw sensor data on local factory edge servers, transmitting only model weight updates to a central aggregator.
- Distribute the current global model state to individual factory edge servers.
- Perform local training on site-specific sensor data to compute weight updates.
- Transmit only the computed weight updates back to the central server.
- Aggregate the received updates using a weighted averaging method to improve the global model.
- Deploy the updated global model back to all factory edge servers for inference.
The cycle of model distribution, local training, and global aggregation.
Local data storage and training
Local data storage and training
Shared intelligence
- Central Server → Distribute weights → Factory Edge A
- Central Server → Distribute weights → Factory Edge B
- Factory Edge A → Send local updates → Central Server
- Factory Edge B → Send local updates → Central Server
- Central Server → Update global state → Global Model
Expected business outcome
A robust, globally optimized predictive maintenance model that respects regional data sovereignty while improving equipment uptime across all locations.
Limitations and checks
Communication overhead remains a constraint; monitor the frequency of synchronization rounds to ensure network stability and check for model convergence degradation.