Federated learning (FL) is a machine learning approach that allows model training across multiple decentralized edge devices (such as smartphones, IoT devices, or local servers) while keeping the data on those devices. It enables model training without the need to centralize sensitive data in a single location, which can be a privacy and security concern.
In FL, the process typically works as follows:
- A global machine learning model is initialized on a centralized server.
- The model is sent to edge devices.
- Edge devices perform local training on their data, making updates to the model using their own data.
- Only the model updates are sent back to the centralized server, not the raw data.
- These updates are aggregated on the server to improve the global model.
- The updated global model is then redistributed to edge devices for further local training.
This iterative process continues, allowing the global model to improve without direct access to the raw data. Federated learning is particularly useful in scenarios where data privacy is a concern, as it limits the exposure of sensitive data while still enabling collaborative model training.
Categories of FL
- Horizontal FL
- Vertical FL
- Federated transfer learning
Open problems
- Improving Efficiency and Effectiveness
- Heterogeneity
- Asymmetric connection speed
- Encryption cost
- Data transmission costs more energy
- Preserving the Privacy of User Data
- Poisoning attack, adversarial attack…
- Differential privacy, Homomorphic Encryption, Secure Multi-Party Computation
- Robustness to Attacks and Failures
- Untargeted attacks, or model downgrade attacks
- Targeted attacks, or backdoor attacks
- Ensuring Fairness and Addressing Sources of Bias
- Bias in Training Data
- Paying for Contributions
- Estimating data owner’s cost incurred for joining the federation
- Estimating data owner’s contribution to the federation