Federated Learning with Adversaries¶
This page is documentation for the FL with Adversaries analysis framework. For now, it provides a reference documentation of all the configuration options that are available in the framework.
Config¶
This is an example of a basic configuration for the framework. This experiment consists of an FL setup with 3383 clients for the Federated-MNIST task. The experiment runs for 80 rounds and in each round, 30 clients are selected The clients train for two epochs and the updates are constrained to have an L2-norm bound of 10.
environment:
num_clients: 3383
num_selected_clients: 30
num_malicious_clients: 0
experiment_name: "Sample run without attackers"
server:
num_rounds: 80
num_test_batches: 5
aggregator:
name: FedAvg
global_learning_rate: -1
client:
clip:
type: l2
value: 10
model_name: resnet18
benign_training:
num_epochs: 2
batch_size: 24
optimizer: Adam
learning_rate: 0.001
dataset:
dataset: femnist
data_distribution: nonIID
Other configuration examples can be found in ./train_configs in the project. For all configuration options, see Framework Configuration