Skip to content

client_selection.md

Latest commit

 

History

History
40 lines (30 loc) · 1.9 KB

client_selection.md

File metadata and controls

40 lines (30 loc) · 1.9 KB

Client Selection

Client selection in Federated Learning (FL) determines which subset of clients participate in each communication round. Intelligent strategies can reduce communication overhead and mitigate non-i.i.d. harms.

List of supported Strategies

Experimental setup

  • Dataset: CIFAR-10
  • Model: ResNet-18
  • Clients: 100 clients total.
  • Data partitioning: Dirichlet distribution with alpha = 0.1 (highly non-IID).
  • Participation per round: 25 clients are sampled each round.
  • Reproduce: 
    python scripts/cs_cifar10_script.py > cs_cifar10_log.txt &
Method Communications Test Loss Accuracy
Uniform 17,767 ± 1,937 0.521 ± 0.009 0.822 ± 0.004
POW-D 10,347 ± 493 0.573 ± 0.012 0.812 ± 0.009
FedCor 19,360 ± 557 0.449 ± 0.017 0.848 ± 0.006
FedCBS 19,207 ± 837 0.507 ± 0.018 0.830 ± 0.007
DELTA 15,700 ± 191 0.816 ± 0.019 0.721 ± 0.007
  • Adversarial client strategies such as POW-D and DELTA provide lower performance while primarily addressing communication overhead.
  • In contrast, FedCor and FedCBS aim to balance performance by selecting the most informative clients, which results in better accuracy but slightly higher communication costs.
  • These differences illustrate the communication-quality trade-off that arises in non-i.i.d. FL.