List your company

Jailbreak Attack Success Rate on VAJA

92.3Identity Attack Success RateZO-SPSA
1.50825.07948.6572.221Jan 5, 2026

Evaluation results

MethodLinks
ZO-SPSAVictim Model=MiniGPT-4, Input=adv. image (ε = 64/255)2026.0192.31008049.6
ZO-SPSAVictim Model=InstructBLIP, Input=adv. image (unconstrained)2026.01809093.880
ZO-SPSAVictim Model=MiniGPT-4, Input=adv. image (unconstrained)2026.0180908060
ZO-SPSAVictim Model=InstructBLIP, Input=adv. image (ε = 64/255)2026.01758081.360
ZO-SPSAVictim Model=LLaVA, Input=adv. image (unconstrained)2026.0170609070
ZO-SPSAVictim Model=InstructBLIP, Input=adv. image (ε = 32/255)2026.01657087.580
ZO-SPSAVictim Model=MiniGPT-4, Input=adv. image (ε = 32/255)2026.0161.571.48050
ZO-SPSAVictim Model=InstructBLIP, Input=adv. image (ε = 16/255)2026.0155607560
ZO-SPSAVictim Model=MiniGPT-4, Input=adv. image (ε = 16/255)2026.0150608050
ZO-SPSAVictim Model=LLaVA, Input=adv. image (ε = 64/255)2026.014066.78050
ZO-SPSAVictim Model=LLaVA, Input=adv. image (ε = 32/255)2026.01366067.545.7
ZO-SPSAVictim Model=LLaVA, Input=adv. image (ε = 16/255)2026.01345562.544.4
No AttackVictim Model=MiniGPT-4, Input=benign image (no attack)2026.0126.248.950.120
No AttackVictim Model=LLaVA, Input=benign image (no attack)2026.0112.24015.444
No AttackVictim Model=InstructBLIP, Input=benign image (no attack)2026.0153012.540