Backdoor Attack on shopping requests GPT-5-generated (test)
94.1Attack Success Rate (ASR)BadDLM
Evaluation Results
| Method | Links | |
|---|---|---|
| BadDLMMethod=BadDLM (Ours), Base Model=Dream-Instruct-7B2026.05 | 94.1 | |
| RL-basedMethod=RL-based, Base Model=Dream-Instruct-7B2026.05 | 69.6 | |
| VPIMethod=VPI, Base Model=Dream-Instruct-7B2026.05 | 48.8 | |
| SFT-basedMethod=SFT-based, Base Model=Dream-Instruct-7B2026.05 | 45.1 | |
| Benign (No Attack)Method=Benign (No Attack), Base Model=Dream-Instruct-7B2026.05 | 5.8 |