Safety Evaluation on MaliciousGen & LMSYS-Chat (test)
97.31Rule ScoreShadow-Level
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Shadow-LevelMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 97.31 | 96.35 | -0.92 | 5.06 | |
| Step-LevelMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 93.08 | 92.12 | -1.05 | 5.16 | |
| Client-LevelMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 83.27 | 73.27 | -1.71 | 5.01 | |
| KrumMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 72.69 | 60.58 | -2.19 | 5.2 | |
| TrimmedMeanMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 68.85 | 48.85 | -2.73 | 5.28 | |
| FedAvgMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 59.42 | 31.35 | -3.49 | 4.93 | |
| ResidualMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 57.69 | 30.77 | -3.5 | 5.08 | |
| FoolsGoldMalicious client ratio=30%, Backbone model=Qwen2.5-7B2026.01 | 55.96 | 23.85 | -3.85 | 5.28 |