In the domain of machine learning (ML), the conventional process typically involves the transfer and storage of data within centralised data centres or on cloud platforms. However, privacy remains a paramount concern due to the inherently sensitive and
confidential nature of most data. As a result, many organisations today struggled with data isolation issues.
To tackle these challenges, a federated learning platform – FedPro, has been developed for better AI training to facilitate cross-institutional and cross-country collaboration. It eliminates the need to transmit raw data, thus preserving
privacy and ensuring compliance with regulatory standards. Furthermore, FedPro empowers users of all skill levels to explore and deploy federated learning applications with little to no coding expertise, making it suitable for a wide array of applications
such as Healthcare, FinTech, Manufacturing, IoT, and Supply Chain.
Features
- Federated learning platform: Offer a user-friendly graphical user interface (GUI) for establishing federated learning and supporting all coordination function
- Easy to use: Allow users to create federations by customising parameters required for model training like the learning rate, local/global epochs, deep learning algorithms, aggregation algorithms and the participant’s datasets
- Decentralised: Support decentralised model aggregation without a central server
- Privacy-preserving: Ensure privacy through differentiable privacy and multi-party computation
- Reliable: Offer real-time monitoring of model training
- Optimised platform: Provide verifiable and immutable storage based on Blockchain
Federated Learning (FL) is an innovative machine learning approach that enables the training of machine learning models using data from different isolated data silos locations worldwide, without the requirement of any data movement.
While most of the current frameworks, platforms, and tools for training FL models are based on Centralised Federated Learning (CFL), in which a central server collects model parameters from multiple clients and executes the aggregate, this centralisation
poses certain challenges such as a single point of failure and communication bottlenecks. FedPro addressed these issue by introducing a novel decentralised model aggregation at numerous clients, which greatly reduces reliance on a single central server
and improving decentralisation. This ground-breaking design could potentially results in significant improvements in scalability, resilience, and overall efficiency.
Managing heterogeneous data distribution from different parties could lead to reduced accuracy and longer convergence times. FedPro develops an advanced weighted-average aggregation method to achieve high performance model aggregation, even when dealing
with skewed data distributions. Additionally, FedPro integrates communication efficient multi-party computation techniques to ensure secure sharing of model parameters during the model aggregation process and implements contribution-aware client selection
to prevent participants from benefitting without contributing their data, all while safeguarding the privacy of the raw data.
Industry Applications
FedPro can be further developed into a suite of IPs in privacy-preserving computing to enhance trust in digital healthcare, and be applied across various industry domain such as Healthcare,
FinTech, Smart Nation and Manufacture, etc.