Bandwidth Slicing to Boost Federated Learning over Passive Optical Networks

During federated learning (FL) process, each client needs to periodically upload local model parameters and download global model parameters to/from the central server, thus requires efficient communications. Meanwhile, passive optical network (PON) is promising to support fog computing where FL tasks can be executed and the traffic generated by FL needs to be transmitted together with other types of traffic for broadband access. In this paper, a bandwidth slicing algorithm in PONs is introduced for efficient FL, in which bandwidth is reserved for the involved ONUs collaboratively and mapped into each polling cycle. Results reveal that the proposed bandwidth slicing significantly improves training efficiency while achieving good learning accuracy for the FL task running over the PON.