Degree-aware In-network Aggregation for Federated Learning with Fog Computing

Data privacy preservation has drawn much attention in emerging machine learning applications, and thus collaborative training is getting much higher such as Federated Learning (FL). However, FL requires a central server to aggregate local models trained by different users. Thus, the central server may become a crucial network bottleneck and limit scalability. To remedy this issue, a novel Fog Computing (FC)-based FL is presented to locally train the model and cooperate to accomplish in-network aggregation to prevent overwhelm the central server. Then, the paper formulates a new optimization problem termed DAT to minimize the total communication cost and maximum latency jointly. We first prove the hardness and propose two efficient algorithms, ADAT-C and ADAT, for the special and general cases, respectively. Simulation and experiment results manifest that our algorithms at least outperform 30% of communication cost compared with other heuristics without sacrificing the convergence rate.