From Bias to Balance: Federated Self-Distillation Retrained Classifier

This project is written in Python for generating dataset distributions and launching federated learning algorithms.

Dataset Generation

First, modify the root_list variable in the load_pick.py file to point to your own dataset paths. For example:

root_list = ["data/cifar10/", 
             "data/cifar100/",
             "data/cinic10/"]

Then run the load_pick.py file to generate the dataset distribution.

Algorithm Startup

Use the following commands to launch the federated learning algorithm:

FedSDC Algorithm

python main.py --config_path /path/to/config/fed_classifier_if01.json --partition_method [if_01|if_001|if_002]

FedSDC-$\omega$ Algorithm

python main.py --config_path /path/to/config/fed_more_contrast_if01.json --partition_method [if_01|if_001|if_002]

Where:

• --config_path specifies the configuration file path
• --partition_method specifies the data partitioning method
  • if_01 indicates an imbalance factor of 10
  • if_001 indicates an imbalance factor of 100
  • if_002 indicates an imbalance factor of 50

With different partition_method parameters, you can obtain different degrees of data imbalance distribution to test the robustness of the algorithm.

Make sure to replace the paths in the commands with your own configuration file paths.