Federated Learning with Generative Adversarial Network to Improve Training on Non-i.i.d Data.

This is the code accompanying the Dissertation Project.
The dissertation report is for application of M.Sc. in Information Technology, The Hong Kong Polytechnic University.

Requirements to run the code:

---

1. Python 3.7
2. PyTorch 1.0.2
3. Scikit-learn
4. Matplotlib
5. Numpy

Important source files:

---

1. FedAVG.py: Implement of the original Federated Average algorithm, which is used as the baseline.
2. Federated_Learning_with_Aug.py: Implement of the FL algorithm with GAN based data augmentation, which is the main function of the project.

Important arguments:

---

The following arguments to the FedAvg_training.py file control the important parameters of the experiment

1. FL_epochs: Defines the number of federated learning training epoch.
2. Gan_epochs: Defines the number of Multi_GAN training epoch.
3. Generator_paths: Number of the path in Generator of the GAN, set to be '4' as a better performance.
4. non_iid_alpha: Hyperparemeter for data partition. Set to be 100 for iid data, and 0.01 for non_iid data.
5. iid: Set for original partition method 'True' or 'False', when data to be partitioned based on the same way in Federated Average algorithm.

Output:

---

The training_loss and test_accuracy is the main output, which can evaluate model performance.
Also, the majority of the information is recorded in the log folder.

Make Contact:

---

If you have questions or suggestions, please make contact in email:smiley:

[email protected]
+852 5480 9259