Design Considerations and Guidelines
leakage . Thus , a fully decentralised or hybrid control is preferred where interaction is peer-topeer among organisations . However , a decentralised consensus over the global model may take more time as compared to centralised aggregation . Moreover , a trustworthy centralised authority may still be in the charge to decide upon the global model architecture , initial hyperparameters setting and debugging .
Implementing fully decentralised orchestration in cross-device FedL involves a huge number of un-reliable clients inching towards the consensus which may result in increased response time . However , managing a huge number of clients by a single orchestrator is also not feasible in terms of monitoring the edge resource information for appropriate client selection . To this end , a local edge server can act as an interface to the global server . The global model can be downloaded beforehand at the local edge servers also gathering the model parameter updates from clients and relaying these to the global server . This hybrid orchestration can unlock the potential of realtime data analytics in true sense by minimizing the prediction delay and maximizing the reliability of clients .
Such hybrid orchestration is realized and tested in one of our previous contributions ( Bharti , 2021 ) in this area . The experiment results showed that utilising edge servers as interaction points for edge devices prevents client failure and minimizes the model convergence time . On the other hand , the model accuracy and convergence time suffers if all the clients are to directly interact with the global server .
As Big Data and deep learning ( Kotsiopoulos , 2021 ) algorithms are key technological enablers for implementing cross-silo FedL , cross-device FedL on the other hand is mainly driven by technologies such as edge computing and 5G to support real-time data analytics .
Figure 4-1 : FedL Implementation Decision Model
28 March 2022