DeceFL to predict leukaemia from A2 benchmark dataset [23]. (A) Data were divided into IID samples for all clients. (B) Data were divided into Non-IID unbalanced samples. (C) Different topologies for FedAvg, SL and DeceFL respectively. The topology for SL must hold in every iteration when any other node is selected as a central client. (D)–(G) Performance of three algorithms on IID/Non-IID setups over logistic regression/neural networks. DeceFL presents a similar performance as FedAvg and SL in both IID and Non-IID scenarios, except during the transient period that DeceFL takes to reach consensus. In contrast to fixed topology in FedAvg and SL, DeceFL provides much more freedom on the choice of communication topology.