List your company

Perturbation-based Jailbreak Attack Defense on Perturbation-based Attack adversarial variants

74.55ASR (PGD, eps=16/255)w/o defense
-1.172418.486338.14557.8037Mar 23, 2026

Evaluation results

MethodLinks
w/o defenseBackbone=Qwen2-VL2026.0374.558076.3777.578078.18
w/o defenseBackbone=MiniGPT-42026.0370.9178.1874.5576.3678.1879.09
w/o defenseBackbone=LLaVA-v1.52026.0369.0974.5580.690.387.2889.09
ASTRABackbone=Qwen2-VL2026.0321.8214.5515.7615.7618.1818.18
NullSteerBackbone=Qwen2-VL2026.0319.9513.8614.2814.3516.8317.28
ASTRABackbone=MiniGPT-42026.035.4512.735.4510.9116.3713.64
NullSteerBackbone=MiniGPT-42026.033.289.434.659.4214.2112.71
ASTRABackbone=LLaVA-v1.52026.031.821.821.210.6100
NullSteerBackbone=LLaVA-v1.52026.031.741.710.6100