Fedlearn

An implementation to simulate a federated learning environment.

It is an attempt to mimic the scenario described in the paper Towards Federated Learning at Scale: System Design

Tech stack

• Python 3.6.1
• Keras
• Flask

Update to previous version

Added a secure aggregator where the model aggregation will be done. The main server can't see the models from the devices. The secure aggregator is a trusted 3rd party where the model aggregation is done.

Run the system using the steps below:

Booting up

1. Run "Device 1" using python app.py
2. Run "Device 2" using python app.py
3. Run "Secure Aggregator" using python sec_agg.py
4. Run "Main Server" using python main_server.py

This will start the Flask servers of the two devices, secure aggregator and the main server.

Servers -

• Main server - http://localhost:8000/
• Device1 - http://localhost:8001/
• Device2 - http://localhost:8002/
• Secure Aggregator - http://localhost:8003/

Everything will work using the REST APIs.

System working

1. First a model will be trained locally on the device. On 'Device1' and 'Device2' server navigate to: http://localhost:8001/modeltrain and http://localhost:8002/modeltrain respectively.

The models will be trained on MNIST data.

2. Once the devices are ready send a status signal to the server that they are ready, using, http://localhost:8001/sendstatus and http://localhost:8002/sendstatus.

There will be a response from the main server.

3. Now, the trained models will be sent to the secure aggregator server using http://localhost:8001/sendmodel and http://localhost:8002/sendmodel

4. The secure aggregator will aggregate the model and build a global model. http://localhost:8003/aggregate_models

5. The secure aggregator will send this aggregated model to the main server. http://localhost:8003/send_model_secagg

6. The main server will send the aggregated model to the devices.

7. http://localhost:8000/send_model_clients

8. Goto step 1. The whole process is repeated again and the aggregated global model is improved at every round.

I tested this on the MNIST data and after 2 rounds got an accuracy of about 95%. The model will eventually update when more devices will be used. The models on devices are being trained on all the train data and also because of this the accuracy is consistent, if the data is partitioned among the devices it would reveal the real performance.

If you would like to improve the current system please feel free to experiment and submit a PR.

Future improvements

• Better model aggregation methods.
• Main server to send parameter configuration to devices based upon their performance.
• A better dataset to simulate the real world usage.
• Android app for simulating devices.
• Main server migration to Heroku or AWS.