Group Collusive Attack Detection on MultiArith
92Detection AccuracyGroupGuard
Evaluation Results
| Method | Links | |
|---|---|---|
| GroupGuardAttack Type=Targeted Suppression, Model=Deepseek-V3.12026.03 | 92 | |
| GroupGuardAttack Type=Targeted Suppression, Model=Qwen3-32B2026.03 | 90 | |
| G-safeguardAttack Type=Targeted Suppression, Model=Deepseek-V3.12026.03 | 89.2 | |
| GroupGuardAttack Type=Group Exclusion, Model=Qwen3-235B-A22B2026.03 | 86 | |
| G-safeguardAttack Type=Targeted Suppression, Model=Qwen3-32B2026.03 | 84.4 | |
| AutoDefenseAttack Type=Group Exclusion, Model=Deepseek-V3.12026.03 | 82.2 | |
| GroupGuardAttack Type=False Consensus, Model=Deepseek-V3.12026.03 | 82 | |
| G-safeguardAttack Type=False Consensus, Model=Qwen3-32B2026.03 | 81.4 | |
| GroupGuardAttack Type=False Consensus, Model=Qwen3-32B2026.03 | 81 | |
| GroupGuardAttack Type=Group Exclusion, Model=Deepseek-V3.12026.03 | 81 | |
| G-safeguardAttack Type=Group Exclusion, Model=Qwen3-235B-A22B2026.03 | 81 | |
| G-safeguardAttack Type=False Consensus, Model=Deepseek-V3.12026.03 | 80.6 | |
| G-safeguardAttack Type=Group Exclusion, Model=Deepseek-V3.12026.03 | 80.5 | |
| AutoDefenseAttack Type=Group Exclusion, Model=Qwen3-235B-A22B2026.03 | 80.2 | |
| G-safeguardAttack Type=Group Exclusion, Model=Qwen3-32B2026.03 | 75.8 | |
| G-safeguardAttack Type=False Consensus, Model=Qwen3-235B-A22B2026.03 | 75.6 | |
| AutoDefenseAttack Type=Group Exclusion, Model=Qwen3-32B2026.03 | 74.6 | |
| GroupGuardAttack Type=Group Exclusion, Model=Qwen3-32B2026.03 | 73 | |
| GroupGuardAttack Type=False Consensus, Model=Qwen3-235B-A22B2026.03 | 71 | |
| G-safeguardAttack Type=Targeted Suppression, Model=Qwen3-235B-A22B2026.03 | 70.2 | |
| AutoDefenseAttack Type=False Consensus, Model=Deepseek-V3.12026.03 | 67.8 | |
| AutoDefenseAttack Type=Targeted Suppression, Model=Qwen3-235B-A22B2026.03 | 66.8 | |
| GroupGuardAttack Type=Targeted Suppression, Model=Qwen3-235B-A22B2026.03 | 66 | |
| AutoDefenseAttack Type=False Consensus, Model=Qwen3-32B2026.03 | 60.6 | |
| AutoDefenseAttack Type=Targeted Suppression, Model=Deepseek-V3.12026.03 | 60.6 | |
| AutoDefenseAttack Type=Targeted Suppression, Model=Qwen3-32B2026.03 | 58.2 | |
| AutoDefenseAttack Type=False Consensus, Model=Qwen3-235B-A22B2026.03 | 54.8 |