Summary
The goal of this chapter was to provide a conceptual overview of the current knowledge of aggregation, the key theoretical step in FL that allows for the disjoint training done by each agent to be pooled together with minimal transmission required. FedAvg is a simple, yet surprisingly powerful aggregation algorithm that performs well in an ideal FL scenario. This scenario is achieved when training is done across IID datasets using machines with similar levels of computational power and no adversarial or otherwise incorrectly performing agents.
Unfortunately, these conditions are often not met when deploying an FL system in the real world. To address these cases, we introduced and implemented modified aggregation approaches: FedProx, FedCurv, and three different robust mean estimators. After reading this chapter, you should have a solid understanding of the considerations that must be taken into account for practical FL applications, and you should be able to integrate the...
