Experimenting Architectural Patterns in Federated Learning Systems

Federated Learning has emerged as a promising paradigm that enables collaborative model training while preserving data privacy, thus contributing to enhance user trust. However, the design of Federated Learning systems requires non-trivial architectural choices to address several challenges, such as system efficiency and learning accuracy. Architectural patterns for Federated Learning systems have been defined in the literature to handle these challenges, but their experimentation is limited. The objective of this paper is to empower software architects in their task of evaluating the design of FL systems while deciding which architectural alternatives are more beneficial in their context of adoption. Our methodology consists of a tool-based approach that embeds the implementation of six architectural patterns defined in the literature. The advantage is that software architects can select design alternatives either in isolation or in a combined fashion, and the subsequent analysis provides the evaluation of some metrics of interest. The experimental results indicate that architectural patterns can enhance system efficiency, although we found a combination of patterns that added overhead and turned to limit its benefit. By quantifying these trade-offs, we aim to support software architects in designing Federated Learning systems by evaluating the benefits and drawbacks of applying architectural patterns.