Experiments are produced on MNIST, CIFAR-10, and WikiText-2 datasets.

• Install the requirements.

pip install -r requirements.txt

• Download the respective datasets and put them under 'data/' directory.

• To run the FedComm experiment:

python fedcomm.py

To run the experiments in different conditions change the parameters in config.json file.

• num_users:Number of total users that have signed up for collaborating. (Default is 100).
• frac: Fraction of users to be used for federated updates. Default is 1.0 (i.e., 100% participation).
• epochs: Number of global training epochs. Default is 1000.
• dataset: Default: 'mnist'. Options: 'mnist', 'cifar10', 'wiki'.

• senders: Fraction of participants in the federated learning scheme that will act as senders. Default: 0.1 (i.e., 10% of the participants).
• payload: The extension of the payload file (under payloads/ directory. Default 'txt', Options: 'txt', 'png'.
• injection: The FL global round when the senders should start transmitting the message. Default 10.
• stealthy: The level of stealthiness of the senders. Default 'non', Options: 'non', 'inter', 'full'.
• run_name: A name given to the particular run. It will create a directory structure where it will store model checkpoints, extracted payloads, train accuracy and loss values.