Robust Safety and Utility Evaluation in Federated Learning on MaliciousGen & WildChat
81.35Rule ScoreStep-Level
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Step-LevelBackbone=Llama3.1-8B, Malicious client ratio=30%, Base Framework=Safe-FedLLM2026.01 | 81.35 | 52.5 | -1.71 | 3.98 | |
| Client-LevelBackbone=Llama3.1-8B, Malicious client ratio=30%, Base Framework=Safe-FedLLM2026.01 | 80 | 55.77 | -1.87 | 4.13 | |
| Shadow-LevelBackbone=Llama3.1-8B, Malicious client ratio=30%, Base Framework=Safe-FedLLM2026.01 | 79.62 | 51.35 | -1.78 | 4.12 | |
| TrimmedMeanBackbone=Llama3.1-8B, Malicious client ratio=30%2026.01 | 48.65 | 5.19 | -3.48 | 3.81 | |
| FedAvgBackbone=Llama3.1-8B, Malicious client ratio=30%2026.01 | 48.08 | 5.77 | -3.59 | 3.82 | |
| ResidualBackbone=Llama3.1-8B, Malicious client ratio=30%2026.01 | 47.31 | 4.81 | -3.49 | 4.12 | |
| FoolsGoldBackbone=Llama3.1-8B, Malicious client ratio=30%2026.01 | 46.73 | 5.96 | -3.48 | 3.72 | |
| KrumBackbone=Llama3.1-8B, Malicious client ratio=30%2026.01 | 46.15 | 5.01 | -3.62 | 3.95 |