CEP

The client eligibility protocol for federated learning was introduced by Asad et al. in 2022. The core idea of this strategy is to select clients based on their performance in 16 4 Implementation previous communication rounds. For this, a subset of clients is initially selected that fulfills the performance requirements. In addition, all clients are given an initial client eligibility score (CES) of 75. Then, based on their previous round performance, they are rewarded or penalized. In our implementation of CEP, each client is rewarded 10 points for successful participation and punished by 5 points if it fails due to a timeout. If a client failed the previous five rounds in a row, they will be penalized by an additional −20 Points. Due to different implementation constraints between the simulation used by Asad et al. and our simulator, we could not implement the full original algorithm. However, our adjusted variant achieves respectable results, in line with those achieved in the original paper.

It is based on the paper

Asad, Muhammad, Safa Otoum, and Saima Shaukat. 2022. 
“Resource and Heterogeneity-Aware Clients Eligibility Protocol in Federated Learning.” 
In GLOBECOM 2022 - 2022 IEEE Global Communications Conference, 1140–45.

The algorithm may be selected by choosing CEP as the algorithm in the config file.

It requires the following parameters:

Key	Description	Example Value
c	Describes the share of all clients to participate in the round	0.2