DeceFL to predict leukaemia from A2 benchmark dataset [23]. (A) Time-varying communication topology that consists of a sequence of graphs each of which is not connected while the lump-sum graph over a fixed period is connected. (B) Time-varying communication topology that adds or removes nodes over time. (C), (D) Performance of DeceFL with edge-varying graphs on the IID and Non-IID setups of dataset A2 using logistic regression, with reference performance of FedAvg that uses full information. (E), (F) Performance of DeceFL with node-varying graphs on the IID and Non-IID setups of dataset A2 using logistic regression, with reference performance of FedAvg that uses full information. These time-varying experiments manifest the robustness of DeceFL in the presence of interventions to communication topology. The dropout or supplement of clients in the middle of DeceFL computation will not break or deteriorate the learning process, which is particularly essential in practice for large-scale time-consuming data-intensive applications.