IEEE Access TDMiL Artifact

This report contains the source code artifact for the IEEE Access'24 paper on TDMiL.

To refer to this work, please cite:

@ARTICLE{10745482,
  author={Gulati, Mayank and Zandberg, Koen and Huang, Zhaolan and Wunder, Gerhard and Adjih, Cedric and Baccelli, Emmanuel},
  journal={IEEE Access}, 
  title={TDMiL: Tiny Distributed Machine Learning for Microcontroller-Based Interconnected Devices}, 
  year={2024},
  volume={12},
  number={},
  pages={167810-167826},
  keywords={Protocols;Training;Internet of Things;Microcontrollers;Machine learning;Data models;Tiny machine learning;Distance learning;Computer aided instruction;Transfer learning;Distributed computing;Federated learning;Tiny machine learning;Distributed learning;federated learning (FL);Internet of Things (IoT);machine learning;microcontrollers;TinyML-as-a-Service (TMLaaS)},
  doi={10.1109/ACCESS.2024.3492921}}

Repository Outline

This repository consists of multiple directories with components.

• RIOT: The directory containing the RIOT source code with minimal modifications
• apps: Demonstration firmware application for the research artifact using RIOT as RTOS.
• coaperator: The CoAPerator orchestrator application written in Python
• modules: Custom modules for the RIOT-based firmware applications
• radvd.d: Default configuration for a router advertisement daemon

How to use this artifact:

Requirements to run the artifacts

RIOT firmware applications

To compile the firmware artifacts, an environment able to compile the RIOT operating system on Linux is required. A Debian-based Linux distribution is recommended. Furthermore the following tools are required:

• Essential system development tools (GNU Make GCC, standard C library headers), can usually be installed by installing build-essential on Debian and derivatives.
• git
• GDB in the multiarch variant (alternatively: install for each architecture you target the corresponding GDB package)
• unzip or p7zip
• wget or curl
• python3
• pyserial (linux distro package often named python3-serial or py3-serial)

These can be installed on Debian and Ubuntu via

# apt install gcc-multilib build-essential git llvm clang python3 python3-pip wget curl unzip

CoAPerator

The CoAPerator orchestrator application written in Python requires at least Python 3.11. The python dependencies can be installed using poetry as detailed in the coaperator directory.

Experiments

Basic experiment

The first experiment launches CoAPerator (the orchestrator) and then a client (TDMiL). The client can then be instructed via CoAPerator to download a model via CoAP. For this, please follow the instructions in apps/FL_ML_client/readme.md

Other experiments

Additional experiments can be found in the apps directory. Head to these directories for instructions on how to run them.