Salvia: Secure Aggregation for Federated Learning in Flower
Status: Completed
Year: 2021
Supervisors: Dr. Nicolas Lane and Dr. Pedro Porto Buarque de Gusmao
Details
This project was completed as part of the Undergraduate Research Opportunities Programme internship at the CaMLSys Lab. Secure Aggregation(SA) protocols in common Federated Learning(FL) frameworks have various limitations, including vulnerability to client dropouts or configuration difficulties. During the internship, I implemented Salvia, a Secure Aggregation feature implemented within the open-source FL framework Flower, that aims to address those limitations:
- Leveraging Flower’s machine learning(ML) framework-agnositc property, Salvia can be used in a FL setting where clients run different ML platforms and pipelines independently.
- Salvia exposes an intuitive and easy-to-use API, allowing uses to freely specify parameters to customize the SA protocol to match their deployment goals.
- Using SecAgg and SecAgg+ as the underlying protocol, Salvia is robust against client dropouts and has asymptotically low theoretical communication and computation overhead.
Results
Experiment results show that Salvia’s performance matches the expected complexities. Future work includes improving the secret sharing functions’ efficiency, and enabling users to customize the weights factor and aggregation computation. A writeup of the project has been accepted to appear in the 2nd International Workshop on Distributed Machine Learning.