PRCSL: A Privacy-preserving Continual Split Learning Framework for Decentralized Medical Diagnosis

This is Pytorch impelementation of "PRCSL: A Privacy-preserving Continual Split Learning Framework for Decentralized Medical Diagnosis", Jungmin Eom, Minjun Kang, Myungkeun Yoon, Nikil Dutt, Jinkyu Kim, Jaekoo Lee

Main Architecture

Dataset

We use MedMNIST, HAM10000, CCH5000, CIFAR100, and SVHN. Datasets should be installed at './src/data'. Before downloading datasets HAM10000 and CCH5000, please refer to ./src/datasets.

• MedMNIST: Run the command "pip install medmnist"
• CIFAR100, SVHN: It is automatically installed when you run the train command with the relevant dataset
• HAM10000
  • Download dataset from https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T
  • To use Data augmentation, refer https://github.com/Woodman718/FixCaps/blob/main/augmentation/02_Aug_img.ipynb
• CCH5000
  • Download dataset from https://zenodo.org/records/53169
  • After download CCH5000, you should convert the format to jpg

Train

Run the following command to train the PRCSL Framework

./scripts/script_cifar100.sh <approach> <gpu> <scenario> [<results_dir>]

The parameters are defined as follows:

• <approach> - approach to be used, from the ones in ./src/approaches/
• <gpu> - index of GPU to run the experiment on
• <scenario> - specific rehearsal scenario
  • base_cl: no exemplars(centralized)
  • fixd_cl: exemplars with fixed memory(centralized)
  • base_csl: no exemplars(split learning)
  • fixd_csl: exemplars with fixed memory(split learning)
  • grow_csl: exemplars with grow memory(split learning)
• [<results_dir>] - results directory (optional), by default it will be ./results

PRCSL code structure is inspired by FACIL.