Krum proposed in: Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent
This example explains how to run the Krum fusion algorithm to train CNNs implemented with Keras training
on MNIST data. Data in this example preprocessed by scaling down to range from [0, 255] to [0, 1].
No other preprocessing is performed.
This experiment can be run using models with different underlying framework. By default, configs with keras(tf 1.15) based model are generated, but other models like PYTORCH, keras(tf 2.1) can be creating by changing -m param.
| Model Type | Params |
|---|---|
| Keras (with tf 1.15) | keras |
| Pytorch | pytorch |
| Tensorflow/keras( tf 2.1) | tf |
Note that the parameter byzantine_threshold is set to 1 by default. According to Krum's assumption, the number of parties in the FL system should be at least 2 * byzantine_threshold + 3. Therefore, to run this example, please create at least 5 parties.
-
Split data by running:
python examples/generate_data.py -n <num_parties> -d mnist -pp <points_per_party> -
Generate config files by running:
python examples/generate_configs.py -f krum -m keras -n <num_parties> -d mnist -p <path> -
In a terminal running an activated IBM FL environment (refer to Quickstart in our website to learn more about how to set up the running environment), start the aggregator by running:
python -m ibmfl.aggregator.aggregator <agg_config>Type
STARTand press enter to start accepting connections -
In a terminal running an activated IBM FL environment, start each party by running:
python -m ibmfl.party.party <party_config>Type
STARTand press enter to start accepting connections.Type
REGISTERand press enter to register the party with the aggregator. -
Finally, start training by entering
TRAINin the aggregator terminal.