FELES is a highly configurable FEderated LEarning Simulator designed for Federated Learning experiments.

Federated Learning

Federated Learning is a new machine learning approach that enables different devices to collaboratively learn a shared prediction model while keeping all the training data on the device.

FL Characteristics

FL has some unique characteristics:

• non-IID data: the training is performed with local data available on devices. Thus data are not independent and identically distributed data (IID)
• heterogeneous network: the devices taking part to the network may be very different in terms of:
  • computation capability
  • network bandwidth and speed
  • energy constraints
• variable network: the property of the devices taking part to the network may change over time, e.g. the network speed available on a device may vary depending on its position
• devices availability: the devices taking part to the network may be available only for a small amount of time
• failures: the devices may fail during the execution of a training or evaluation job due to different reasons, e.g. unavailable network or available resources
• high number of devices: a high number of devices may be used to train a model

FL Experiments Limits

Empirical evaluations of FL approaches may be difficult to be replicated and compared:

• FL network deployment may be complex and unfeasible
• experiment cost may be very high due to high number of devices
• the experiment may be difficult to be replicated
• lack of standardized benchmarks
• lack of standardized frameworks

FELES Goals

FELES provides the following features:

• orchestrator / worker architecture: the experiments can be executed using only a single machine avoiding the need for an expensive FL network. Multiple workers can be used to scale the simulation
• hardware independent: the results do not depend on the hardware where the experiment was executed
• reproducibility: the experiments are easily reproducible with a configurable environment
• development and comparison: it facilitates the development, test, comparison and analysis of custom algorithms with popular ones (e.g. FedAvg, FedProx)
• customizable: it is easily customizable allowing to plug-in optimizers, logic blocks defining the strategy applied at different stages of the FL protocol, e.g. client selection)
• results analyzer: the simulator provide an analyzer for plotting and visualizing the resulting data

Getting Started

Setup

virtualenv env
pip install -r requirements.txt

Start a Simulation

1. the simulation parameters are defined in the config.py file. A complete explanation of the available parameters is given in Parameters
2. open a shell and start the orchestrator with python simulate_orchestrator.py
3. open another shell and start the worker with python simulate_worker.py
4. the output of the simulation will be saved in the output folder by default

Analyze the Simulation data

The Analyzer can be started with python analyze.py -f <list of files> -d -p

• the files after -f are read from the output folder
• -d print the simulation data
• -p export the plots

Documentation

More information about parameters, optimizers and the analyzer are available here