Fig 1.
The overview of system model.
Table 1.
Comparison with conventional hyperparameter search.
Table 2.
Summary of experimental datasets and partitioning strategies.
Table 3.
Experimental setup.
Table 4.
Hyperparameter settings for all compared methods.
Fig 2.
IID settings for different datasets.
Fig 3.
Non-IID settings for different datasets.
Table 5.
Function comparison of different schemes.
Fig 4.
Comparison of different schemes (CIFAR-10 dataset).
Fig 5.
Comparison of different schemes (MNIST dataset).
Table 6.
Sensitivity analysis of gradient sparsification ratio k on CIFAR-10, VGG11, 20 rounds.
Fig 6.
(Edge device computing overhead) comparison of different schemes.
Fig 7.
(Cloud server computing cost) comparison of different schemes.
Table 7.
Trade-off between decision cost and communication benefit.
Table 8.
System-level profiling — CIFAR-10 / VGG11 / C = 10 / 10 Mbps uplink.
Table 9.
System-level profiling — MNIST / 2NN / C = 10 / 10 Mbps uplink.
Table 10.
Ablation study results — CIFAR-10 dataset (C = 10, 20 rounds, VGG11).
Table 11.
Ablation study results — MNIST dataset (C = 10, 20 rounds, 2NN).
Table 12.
Comparison of fixed-rate pruning and APMKFL on CIFAR-10 and MNIST datasets.