List your company
ResearchFormerly WizWand

Multimodal Prompt Injection Attack on ImageNet

0Attack Success Rate (ASR)ATPI
-3.78221.746547.27572.8035May 15, 2026
Updated 5mo ago

Evaluation results

MethodLinks
ATPITarget Model=InstructBLIP, Best Source Model=MiniGPT4-llama22026.05028.62
ATPITarget Model=BLIP-2, Best Source Model=MiniGPT4-llama22026.05013.29
ARE-WTarget Model=MiniGPT4-llama2, Best Source Model=MiniGPT4-vicuna2026.052.0428.61
ARE-WTarget Model=BLIVA, Best Source Model=MiniGPT4-llama22026.054.0820.27
ATPITarget Model=BLIVA, Best Source Model=MiniGPT4-llama22026.055.832.88
ARE-WTarget Model=InstructBLIP, Best Source Model=MiniGPT4-llama22026.056.1220.84
ARE-WTarget Model=BLIP-2, Best Source Model=MiniGPT4-llama22026.058.1619.71
ARE-WTarget Model=MiniGPT4-vicuna, Best Source Model=MiniGPT4-llama22026.0510.229.32
ATPITarget Model=MiniGPT4-llama2, Best Source Model=MiniGPT4-vicuna2026.0510.5332.28
ATPITarget Model=Qwen2.5-VL, Best Source Model=Qwen2.5-VL2026.0512.231.11
ARE-WTarget Model=Qwen2.5-VL, Best Source Model=Qwen2.5-VL2026.0513.9518.89
ATPITarget Model=MiniGPT4-vicuna, Best Source Model=MiniGPT4-llama22026.0517.5429.92
CITarget Model=MiniGPT4-llama2, Best Source Model=MiniGPT4-vicuna2026.0520.3722.44
ARE-BTarget Model=MiniGPT4-llama2, Best Source Model=MiniGPT4-vicuna2026.0522.7332.41
CrossMPITarget Model=MiniGPT4-llama2, Best Source Model=MiniGPT4-vicuna2026.0527.1832.63
CrossMPITarget Model=Qwen2.5-VL, Best Source Model=Qwen2.5-VL2026.0529.4121.87
ARE-BTarget Model=Qwen2.5-VL, Best Source Model=Qwen2.5-VL2026.0536.3622.56
CITarget Model=MiniGPT4-vicuna, Best Source Model=MiniGPT4-llama22026.0545.4533.18
ARE-BTarget Model=InstructBLIP, Best Source Model=MiniGPT4-llama22026.0547.7330.76
ARE-BTarget Model=BLIVA, Best Source Model=MiniGPT4-llama22026.0547.7328.18
ARE-BTarget Model=MiniGPT4-vicuna, Best Source Model=MiniGPT4-llama22026.0556.8242.89
ARE-BTarget Model=BLIP-2, Best Source Model=MiniGPT4-llama22026.0556.8241.76
CITarget Model=Qwen2.5-VL, Best Source Model=Qwen2.5-VL2026.0567.2729.49
CrossMPITarget Model=MiniGPT4-vicuna, Best Source Model=MiniGPT4-llama22026.0569.951.01
CITarget Model=BLIVA, Best Source Model=MiniGPT4-llama22026.0579.2137.33
CITarget Model=InstructBLIP, Best Source Model=MiniGPT4-llama22026.0581.8236.46
CrossMPITarget Model=BLIP-2, Best Source Model=MiniGPT4-llama22026.0593.254.17
CrossMPITarget Model=InstructBLIP, Best Source Model=MiniGPT4-llama22026.0594.1737.56
CrossMPITarget Model=BLIVA, Best Source Model=MiniGPT4-llama22026.0594.1736.08
CITarget Model=BLIP-2, Best Source Model=MiniGPT4-llama22026.0594.5550.17