FL in the AI4EOSC platform (demo)
FL in the AI4EOSC platform (demo) - Judith Sáinz-Pardo Díaz (CSIC)
Services for AI/ML development
Training: federated learning
- Collaborative and decentralized approach to build ML models.
- No need to centralize a dataset.
- Management of experiments through platform dashboard.
- Participating clients both within AI4EOSC platform or external (with authentication).
Client-server authentication
- Simplest configuration: clients only need to enter the deployment ID where the server is deployed to join the federated learning training.
- An authentication system (using tokens) for clients prior to their incorporation into the federated training has been implemented.
- Vault has been adopted for storing the tokens.
- The FL server retrieves a list of authorized tokens for client authentication from Vault and checks the token provided by the clients.
- Users can manage the tokens (add or revoke).
- Some extensions to the Flower library to work alongside Vault have been implemented.
- Some modifications to the Flower library to manage the credentials have been performed.
Federated learning in AI4EOSC
GOAL: classify chest X-Ray images according to whether or not the patient has pneumonia.
We divide the initial train data into 3 clients. Stratified train-test random split: 75% train, 25% test.
Model: multi-layer convolutional network implemented using keras.
This part of the course is very practical so most of the information comes from watching the video.