ActiveFL

ActiveFL is an algorithm introduced in 2019 with the goal of selecting an optimized subset of clients to participate in federated learning based on a value function. This value function aims to quantify the expected usefulness of a client’s data during a training round. In addition to sampling a subset of the clients based on their valuation, another set is selected uniformly at random. While not explicitly mentioned in the original paper we assume that the random portion of the sampled set aims to reduce overfitting on the same clients every round.

It is based on the paper

Goetz, Jack, Kshitiz Malik, D. Bui, Seungwhan Moon, Honglei Liu, and Anuj Kumar. 2019. 
“Active Federated Learning.” arXiv.org. 
https://www.semanticscholar.org/paper/36b9b82b607149f160abde58db77149c6de58c01.

The algorithm may be selected by choosing ActiveFL 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
alpha1	Percentage of lowest score clients to disqualify	0.75
alpha2	Parameter for selection probability	0.01
alpha3	Percentage of clients to select randomly	0.1