Towards A Trustworthy Federated Learning Framework for Constrained Devices

Federated learning (FL) is a relatively new approach to
machine learning that aims to ensure data privacy is preserved by
training models on a distributed network of clients. While FL has
shown potential, challenges such as inefficient communication schemes
and poor performance with non-independent and identically distributed
data pose significant barriers to its wider adoption. This work proposes
a framework to overcome some of these challenges by exploring
methods to improve communication reliability, improve client selection
and investigate pathways to implement robust trust mechanisms. The
goal is to develop a trustworthy FL framework with a focus on
constrained IoT devices that will be applicable to a range of use cases
such as those found in industrial settings, where there is a need to
improve efficiency and safety, or detecting fraudulent transactions on
mobile devices, helping protect customers. A FL network of this type
must be capable of handling devices that produce poor quality data,
have an efficient communication scheme, and efficient training processes
to preserve battery life, while working efficiently with multi-modal
data. This paper will first provide background information on FL,
discuss the approach to address some of its challenges, and detail the
framework’s verification process. Additionally, it will present initial
baseline experiments using the Flower framework to develop an
understanding of how FL networks are created and implemented, and
to assess the impact of varying the number of clients or training rounds
on overall accuracy.