Experiments

1) Main experiments

We choose baseline hyperparameters and vary one at a time.

Possible baseline

• Aggregation: FedAVG
• Clients: 40
• Communication rounds: 20
• Clients sampled: 20
• Local data amount: 1250
• Local epochs: 20
• alpha: 1
• Batch size: 32
• No noisy clients
• CIFAR10

Experiments

• Varying number of clients sampled
  • 5, 10, 20, 40
  • 5x
• Varying local epochs
  • 1, 10, 20, 40, 80
  • 1x på Pi med maks-tid til tidsplot
  • 5x på HPC til comm round-plot
• Varying alpha + IID
  • IID, 10^-2, 1, 10^2
  • 5x
• Varying noisy clients
  • 0, 10, 20, 30, 40
  • 5x
• FedDF: Repeat alpha and noise experiments

2) Comparison to centralized learning

• Run baseline for "unlimited" communication rounds to show that after enough time, both FedAVG and FedDF can reach centralized