Federated Learning is a decentralized machine learning paradigm in which a multitude of clients locally train models on their private data and only transmit model parameters or gradients to a central aggregator.
The aggregator typically combines the updates using weighted averaging (e.g., Federated Averaging, FedAvg) to create a global model, which is then sent back to the clients.
This procedure enables training models on non-identically distributed (non-IID) and heterogeneous data without transferring the data itself, thereby improving privacy and data security.
Key research questions include optimizing communication costs, robustness against faulty or malicious clients, and ensuring privacy through techniques such as Differential Privacy and Secure Multi-Party Computation.
Definition:
Federated Learning is a machine learning method in which multiple decentralized clients train locally and only send their model updates to a central server that aggregates them to create a shared model without the raw data leaving the clients.
Source:
Konecny, J., McMahan, H. B., Yu, F., Richtárik, P., Suresh, A. T., & Bacon, D. (2016). Federated Learning: Strategies for Improving Communication Efficiency. arXiv preprint arXiv:1610.05492.