This repository contains the code for the Android implementation of the HgO algorithm. This Application is for demonstration purposes and it is by no means built for production usage.
The backend folder contains the code for the parameter server (PS). It can be deployed on a local machine or in the cloud with a public IP address.
The PS accepts connections and disconnections from any device at any time. Each connected device receives the machine learning model with current parameters. Once a new round is started, the PS sends the model parameters and the selected block to be updated in the next round.
A disconnected or a stale device is considered as a new device, and receives the most recent model parameters once it is connected again.
pip install -r requirements.txt
To configure the Parameter Server, update the configuration file backend/config.py. The default parameters are:
# file: backend/config.py
SERVER_HOST = "0.0.0.0"
SERVER_PORT = 45000
ROUNDS = 500
FRAC = 0.8
BLOCK_SIZE = 0
LEARNING_RATE = 0.001
GAR = "average"
F = 0
ATTACK = "No"
MIN_ACTIVE_WORKERS = 2
Download the MNIST dataset for model evaluation and move it to the datasets folder backend/datasets.
MNIST download link: http://tiny.cc/mnist
To run the Parameter Server, execute the following commands:
cd backend
python main.py
The smartphone implementation of HgO uses the python cross platform library kivy (\url{https://kivy.org}). Our code is optimized to deploy the Android version. However, the application is cross platform and can also be deployed on IOS, Linux, macOS and Windows, with the appropriate modifications. The application works on Android version 7+ and only requires storage permission to access the smartphone local dataset.
Install buildozer: https://buildozer.readthedocs.io/en/latest/installation.html
Execute the following command to build the Android application:
cd frontend
buildozer android debug deploy run logcat
A built application is available in the frontend/dist/ folder.
The application needs first to be configured by setting the correct IP address and port of the server. Next, the computation profile needs to be selected. For simplicity, we suggest three configurable computation profiles with the following default values:
- Low configuration: train the model using only 16 data points at each round.
- Moderate configuration: train the model using 128 data points at each round.
- Powerful configuration: train the model using the number of data points chosen specified in the “Number of samples” text input.
Once configured, the application will receive the selected model from the PS and start training once the PS sends the current model parameters and the selected block to train in the next round. Once the training is finished, a new screen is activated with a summary of the training.
