Modifying aggregation for non-ideal cases
In practical FL applications, at least one of the aforementioned assumptions that constitute an ideal FL scenario generally does not hold; therefore, the usage of alternative aggregation methods might be necessary to best perform FL. The goal of this section is to cover examples of aggregation methods that target heterogeneous computational power, adversarial agents, and non-IID datasets, in order of difficulty.
Handling heterogeneous computational power
As mentioned earlier, the ideal aggregation approach, in this case, consistently avoids the straggler effect while maximizing the number of agents participating in FL and allowing all agents to contribute to some extent, regardless of computational power differences. Agents become stragglers during a round when their local training takes significantly more time than the majority of the agents. Therefore, effectively addressing this problem actually requires some level of adaptability...
