Image Classification on CIFAR-10 (test) (Specific Adversarial Attack Robustness)
91.2Accuracy (No Attack)Cost-TrustFL
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Cost-TrustFLMalicious client ratio=30%, Dirichlet alpha (α)=0.5, Backbone=CNN (2 conv, 2 fc layers), Communication rounds=200, Local epochs (E)=5, Batch size=32, Learning rate=0.01, Reference dataset size=100 samples, Cost weight (λ)=0.32025.12 | 91.2 | 86.7 | 87.8 | 85.5 | 84.1 | |
| FLTrustMalicious client ratio=30%, Dirichlet alpha (α)=0.5, Backbone=CNN (2 conv, 2 fc layers), Communication rounds=200, Local epochs (E)=5, Batch size=32, Learning rate=0.01, Reference dataset size=100 samples2025.12 | 90.5 | 83.1 | 84.5 | 81.2 | 79.2 | |
| Trimmed-MeanMalicious client ratio=30%, Dirichlet alpha (α)=0.5, Backbone=CNN (2 conv, 2 fc layers), Communication rounds=200, Local epochs (E)=5, Batch size=32, Learning rate=0.012025.12 | 89.8 | 78.9 | 79.8 | 74.5 | 71.2 | |
| FedAvgMalicious client ratio=30%, Dirichlet alpha (α)=0.5, Backbone=CNN (2 conv, 2 fc layers), Communication rounds=200, Local epochs (E)=5, Batch size=32, Learning rate=0.012025.12 | 89.1 | 68.3 | 54.5 | 41.2 | 32.8 | |
| KrumMalicious client ratio=30%, Dirichlet alpha (α)=0.5, Backbone=CNN (2 conv, 2 fc layers), Communication rounds=200, Local epochs (E)=5, Batch size=32, Learning rate=0.012025.12 | 88.5 | 76.1 | 77.2 | 69.8 | 64.5 |