Learning with real-world sensitive data (e.g. healthcare records, biomedical and financial data) is facing a big challenge due to privacy concerns and legislation such as General Data Protection Regulation (GDPR).
Recent advances in distributed learning methods of Federated Learning (FL) have shown promise in training models, particularly in the areas of collaborative working over geographically distributed computing nodes without sharing individual private data, keeping both data and computation locally and aggregating locally computed results on a server to train a global model. The general concept of FL is to use a form of distributed stochastic gradient descent (SGD) and requires a parameter server to coordinate the training process, which involves gradients computation and aggregations.
This poses several open challenges during the federated learning process:
1) how to reduce communication costs when transferring computation results between working nodes and a server? 2) how to ensure the SGD is an unbiased estimation of the full gradient with those non-identically independently distributed samples (N-IID)? 3) how to optimise model parameters for computing efficiency?
To address the challenges above, this project will develop a novel efficient privacy preserving federated learning framework over decentralised data, which will be validated with a real world application scenario.
Aims and objectives
This project aims to develop a novel efficient distributed learning framework over decentralised data while preserving privacy. The model will be tested and evaluated with real data sets in the healthcare domain.
The objectives are: 1.To conduct a comprehensive literature review on distributed machine learning/deep learning/big data analytics; 2.To develop an efficient strategy to reduce communication costs when transferring local computation results (e.g. weights of deep networks) between a distributed node and a parameter server; 3.To develop an efficient federated learning model which can handle Non-IDD issue; 4.To optimise the model for efficiency; 5.To test and evaluate the model with real data sets.
Specific requirements of the project
The candidate is expected to possess the following academic qualifications and skills. 1.1st or 2:1 honours degree above in related areas (Computer Science) or MSc. degree (Merit). 2.Excellent analytical skills/machine learning or deep learning/programming skills. 3.Good team working skills. 4.Experience of having worked with industrial partners is desirable.
This opportunity is open to UK, EU and overseas students. Funding is available to the equivalent of Home/EU fees - overseas applicants will need to pay the difference in fees.