Privacy-Preserving Collaborative Learning for Diabetic Retinopathy

Research Intern: Eleanor Cooper 

Project supervisor: Steven Kerr 

PhD Tutor: Bryan Lopez De Munain 

In this project I explored whether separate retinal image datasets can collaboratively contribute to training LLM without pooling the raw patient images and data. I used RETFoundGreen (RFG) and explored to approaches: Federated Learning (FL) and Secure multi-party computation (SMPC) 

Due to infeasible SMPC costs, federated learning was used to adapt the RFG encoder while keeping the raw images local to two clients representing my two datasets: IDRiD and MESSIDOR-2. First, FL was manually implemented to confirm validity before the FlowerAI framework was used. I then evaluated federated and centralised adapters under a controlled downstream classifier.  The results demonstrated that federated adaptation of the encoder was feasible, although performance varied between the two datasets and the experiments were limited by a small training budget.  

I then investigated SMPC as a stronger cryptographic approach in which parties jointly compute over secret-shared values. I used MP-SPDZ framework and fixed 384-D RFG embeddings and them implemented and validated secure training against matched pooled calculations.  Secure logistic regression demonstrated the feasibility of the approach but required approximately 5.1 hours and 5.48 TB of communication so i instead investigate alternatives like polynomial kernel ridge regression and ridge regression. Ridge regression proved favourable as it produced  a combined AUROC of 0.8620, only 0.0043 below the scikit-learn reference, while reducing runtime to 19.7 minutes and global communication to 359 GB.  

Overall, this project showcased that both approaches are feasible however that thought must be put into the mathematical formulation used in order to ensure practicality and accuracy.