Toxicity Generation on RealToxicityPrompts (test)
9.2Perspective API ScoreWithout Attack
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Without AttackTarget Model=LLaVA, Target Backbone=LLaMA-2-Chat2026.01 | 9.2 | 6.4 | |
| Without AttackTarget Model=InstructBLIP, Target Backbone=Vicuna2026.01 | 34.2 | 36.4 | |
| Without AttackTarget Model=MiniGPT-4, Target Backbone=Vicuna2026.01 | 34.8 | 29.1 | |
| ZO-SPSASurrogate Model=InstructBLIP, Surrogate Backbone=Vicuna, Target Model=MiniGPT-4, Target Backbone=Vicuna2026.01 | 46.67 | 38.34 | |
| ZO-SPSASurrogate Model=LLaVA, Surrogate Backbone=LLaMA-2-Chat, Target Model=MiniGPT-4, Target Backbone=Vicuna2026.01 | 47.84 | 39.76 | |
| ZO-SPSASurrogate Model=MiniGPT-4, Surrogate Backbone=Vicuna, Target Model=MiniGPT-4, Target Backbone=Vicuna2026.01 | 54.11 | 61.05 | |
| ZO-SPSASurrogate Model=LLaVA, Surrogate Backbone=LLaMA-2-Chat, Target Model=LLaVA, Target Backbone=LLaMA-2-Chat2026.01 | 61.4 | 61.4 | |
| ZO-SPSASurrogate Model=MiniGPT-4, Surrogate Backbone=Vicuna, Target Model=LLaVA, Target Backbone=LLaMA-2-Chat2026.01 | 61.52 | 54.03 | |
| ZO-SPSASurrogate Model=InstructBLIP, Surrogate Backbone=Vicuna, Target Model=LLaVA, Target Backbone=LLaMA-2-Chat2026.01 | 61.58 | 53.76 | |
| ZO-SPSASurrogate Model=LLaVA, Surrogate Backbone=LLaMA-2-Chat, Target Model=InstructBLIP, Target Backbone=Vicuna2026.01 | 63.13 | 54.98 | |
| ZO-SPSASurrogate Model=InstructBLIP, Surrogate Backbone=Vicuna, Target Model=InstructBLIP, Target Backbone=Vicuna2026.01 | 63.95 | 61.18 | |
| ZO-SPSASurrogate Model=MiniGPT-4, Surrogate Backbone=Vicuna, Target Model=InstructBLIP, Target Backbone=Vicuna2026.01 | 64.18 | 45.01 |