BlockFLow Supplementary materials

Contents

Our supplementary materials submission contains the following:

1. Paper with Appendix
2. Data for all experiments in the results folder
3. Code to reproduce figures that appear in the paper in the experiments_and_figures folder
4. Source code in the [src] folder
5. Intructions to reproduce results, below:

Result Reproduction Instructions

1. Install Python 3.8 on Linux. We have not tested BlockFLow on other platforms.
2. Run pip install -r src/requirements.txt
3. Run all cells in src/examples/logistic_regression/adult/generate_datasets.ipynb, specifying paths as appropriate
4. Run all cells in src/examples/logistic_regression/kdd99/generate_datasets.ipynb, specifying paths as appropriate
5. Install Docker
6. Results in experiments 1-4 can be reprocduced through off-chain simulation of BlockFLow. To do so,
   1. Run all cells in experiments_and_figures/exp1_thru_4.ipynb, specifying paths as appropriate
   2. For experiments 1, 2, or 4:
      1. Run all cells in experiments_and_figures/exp{NUMBER}.ipynb, where NUMBER is 1, 2, or 4. Specify paths as appropriate
   3. For experiment 3:
      1. Run all cells in experiments_and_figures/score_exp3.ipynb, specifying paths as appropriate
      2. Run all cells in experiments_and_figures/exp3.ipynb, specifying paths as appropriate
7. Results in experiment 5 can only be reporduced through an on-chain simulation of BlockFLow. To do so,
   1. Install Docker Compose
   2. Copy src/examples/example.config.py to src/examples/config.py. Set RESULTS_FOLDER_PATH to some empty folder
   3. Run all cells in src/infra/geth/generate_gensis.ipynb
   4. Copy the JSON-stringification of genesis to src/infra/geth/genesis.json
   5. Copy the value of the private_keys variable to the value of PRIVATE_KEYS in src/examples/config.py
   6. From src/infra/local, run docker-compose up --build
   7. From src, run python3 -m examples.logistic_regression.run_multi_clients --exp_name adult --dataset_folder /path/to/data/root/adult/split_sym_eq_1_validation_fraction_0.2 --ground_truth_dataset /path/to/data/root/adult/test.dat
   8. From src, run python3 -m examples.logistic_regression.run_multi_clients --exp_name adult --dataset_folder /path/to/data/root/adult/split_sym_eq_3_validation_fraction_0.2 --ground_truth_dataset /path/to/data/root/adult/test.dat
   9. From src, run python3 -m examples.logistic_regression.run_multi_clients --exp_name adult --dataset_folder /path/to/data/root/adult/split_sym_eq_5_validation_fraction_0.2 --ground_truth_dataset /path/to/data/root/adult/test.dat
   10. Inside RESULTS_FOLDER_PATH/tensorboard, run tensorboard --logdir .
   11. Tensorboard visualizes the gas consumption by federated learning round and provides links to export such data into CSV or JSON format. Copy data into statistical software to perform linear regressions.