Model-Heterogeneous-FL-Attacks

Requirements:

Refer to requirements.txt. To install necessary packages, run pip install -r requirements.txt

Code Structure:

• train_classifier_fed.py: High level source file for the Convergence Rate Attack.
• train_classifier_rolex.py: High level source file for the Rolling Model Attack.
• fed.py: Controls submodel distribution/aggregation for the Convergence Rate Attack.
• fed_rolex.py: Controls submodel distribution/aggregation for the Rolling Model Attack.

Parameters:

• Hyperparameters are set through the config.yml and utils.py files.
• To specify an output directory for the Tables, change cfg['file_output'] for the attack you are running (defined in train_classifier_fed.py and train_classifier_rolex.py).
• To alter the local training dataset size, change cfg['local_train_size']
• For Rolling Model Attack:
  • Change the noise percentage with cfg['noise_scale']
  • Change the initial distribution value with cfg['distribute_init_val']

How to Run:

Navigate to source directory. Then, run the desired attack using one of the commands below.

Convergence Rate Attack: python3 train_classifier_fed.py --data_name MNIST --model_name fcnn --control_name 1_5_1_non-iid-2_fix_a1-b1-c1_bn_1_1

Rolling Model Attack: python3 train_classifier_rolex.py --data_name MNIST --model_name fcnn --control_name 1_3_1_iid_fix_a1-b1-c1_bn_1_1

Output:

• The table file will summarize the relevant attack metrics.