Federated Learning

Status: public · Confidence: medium (0.815) · Basis: verified_sources

## TL;DR

Federated learning trains a shared model while data stays distributed across clients. It reduces the need to centralize raw data, but privacy is a protocol property, not a slogan: aggregation, update security, differential privacy, and deployment governance still matter.

## Core Explanation

The basic loop is simple. A server sends a model to clients, clients train locally on their own data, and the server aggregates model updates into a new global model. FedAvg is the canonical version of this loop: multiple local gradient steps happen on clients before their model updates are averaged.

The practical difficulty is that clients are not identical. Devices may be offline, slow, or resource-limited, and local data is often non-IID. Secure aggregation and privacy accounting are separate layers that can reduce leakage from updates.

## Further Reading

- [Federated Averaging](https://arxiv.org/abs/1602.05629)
- [Advances and Open Problems in Federated Learning](https://arxiv.org/abs/1912.04977)
- [Practical Secure Aggregation](https://arxiv.org/abs/1611.04482)

## Related Articles

- [AI for IoT](./ai-for-iot.md)
- [Adversarial Machine Learning](./adversarial-machine-learning.md)
- [Public-Key Cryptography](../computer-science/public-key-cryptography.md)