List your company

Black-box Attack on VLM Evaluation Set (test)

75.75Ensemble Success RateSGHA-Attack (Ours)
29.958841.846953.73565.6231Feb 2, 2026

Evaluation results

MethodLinks
SGHA-Attack (Ours)VLM Model=BLIP-2, Defense=Bit Reduction2026.0275.7591.973.9
SGHA-Attack (Ours)VLM Model=InstructBLIP, Defense=Bit Reduction2026.0274.8493.374.4
SGHA-Attack (Ours)VLM Model=UniDiffuser, Defense=Bit Reduction2026.0274.5297.381.7
SGHA-Attack (Ours)VLM Model=UniDiffuser, Defense=JPEG Compression2026.0272.2597.177.1
COAVLM Model=BLIP-2, Defense=Bit Reduction2026.0269.1290.556.4
COAVLM Model=UniDiffuser, Defense=Bit Reduction2026.0268.3398.465.6
COAVLM Model=InstructBLIP, Defense=Bit Reduction2026.0268.392.256
MF-iiVLM Model=UniDiffuser, Defense=Bit Reduction2026.0265.9994.460
COAVLM Model=UniDiffuser, Defense=JPEG Compression2026.0265.5198.456.8
MF-iiVLM Model=BLIP-2, Defense=Bit Reduction2026.0265.3275.542.6
SGHA-Attack (Ours)VLM Model=UniDiffuser, Defense=ComDefend2026.0265.059156
MF-iiVLM Model=InstructBLIP, Defense=Bit Reduction2026.0264.127743.5
SGHA-Attack (Ours)VLM Model=LLaVA, Defense=Bit Reduction2026.0264.0293.576.4
MF-iiVLM Model=UniDiffuser, Defense=JPEG Compression2026.0263.1691.351.2
SGHA-Attack (Ours)VLM Model=BLIP-2, Defense=JPEG Compression2026.0258.8256.427.4
COAVLM Model=UniDiffuser, Defense=ComDefend2026.0258.2393.334.5
SGHA-Attack (Ours)VLM Model=InstructBLIP, Defense=ComDefend2026.0257.6651.723.3
SGHA-Attack (Ours)VLM Model=InstructBLIP, Defense=JPEG Compression2026.0257.475425.1
SGHA-Attack (Ours)VLM Model=BLIP-2, Defense=ComDefend2026.0257.3851.824.3
SGHA-Attack (Ours)VLM Model=UniDiffuser, Defense=DiffPure2026.0256.8670.127.2
AdvDiffVLMVLM Model=BLIP-2, Defense=Bit Reduction2026.0256.7545.919.6
AdvDiffVLMVLM Model=InstructBLIP, Defense=Bit Reduction2026.0256.4548.719.3
MF-iiVLM Model=UniDiffuser, Defense=ComDefend2026.0256.1179.128.7
SGHA-Attack (Ours)VLM Model=InstructBLIP, Defense=DiffPure2026.0256.144.918.2
SGHA-Attack (Ours)VLM Model=BLIP-2, Defense=DiffPure2026.025643.318.5
SGHA-Attack (Ours)VLM Model=MiniGPT-4, Defense=Bit Reduction2026.0255.7493.973.4
AdvDiffVLMVLM Model=BLIP-2, Defense=DiffPure2026.0255.1241.513.7
AdvDiffVLMVLM Model=UniDiffuser, Defense=JPEG Compression2026.0254.9772.224
AdvDiffVLMVLM Model=InstructBLIP, Defense=DiffPure2026.0254.534315
COAVLM Model=UniDiffuser, Defense=DiffPure2026.0254.4783.420.3
COAVLM Model=BLIP-2, Defense=JPEG Compression2026.0254.4660.215.6
AdvDiffVLMVLM Model=UniDiffuser, Defense=Bit Reduction2026.0254.2269.221.6
COAVLM Model=InstructBLIP, Defense=ComDefend2026.0254.255.514.6
COAVLM Model=BLIP-2, Defense=ComDefend2026.0254.1355.614.4
COAVLM Model=InstructBLIP, Defense=JPEG Compression2026.0254.0359.315.7
SGHA-Attack (Ours)VLM Model=LLaVA, Defense=JPEG Compression2026.0253.870.138.1
COAVLM Model=BLIP-2, Defense=DiffPure2026.0253.374811.6
COAVLM Model=InstructBLIP, Defense=DiffPure2026.0253.1847.811.6
AdvDiffVLMVLM Model=BLIP-2, Defense=JPEG Compression2026.0252.8142.511.9
AdvDiffVLMVLM Model=InstructBLIP, Defense=ComDefend2026.0251.7334.38.9
MF-iiVLM Model=BLIP-2, Defense=JPEG Compression2026.0251.6937.412.4
MF-iiVLM Model=InstructBLIP, Defense=ComDefend2026.0251.6132.89.7
MF-iiVLM Model=LLaVA, Defense=Bit Reduction2026.0251.666.230.3
MF-iiVLM Model=BLIP-2, Defense=DiffPure2026.0251.5827.47.1
MF-iiVLM Model=InstructBLIP, Defense=DiffPure2026.0251.2627.47.6
AdvDiffVLMVLM Model=InstructBLIP, Defense=JPEG Compression2026.0251.2338.711.5
AdvDiffVLMVLM Model=UniDiffuser, Defense=DiffPure2026.0251.162.814.4
AdvDiffVLMVLM Model=BLIP-2, Defense=ComDefend2026.0250.9832.38.1
MF-iiVLM Model=InstructBLIP, Defense=JPEG Compression2026.0250.9833.710.6
MF-iiVLM Model=BLIP-2, Defense=ComDefend2026.0250.9531.49
MF-iiVLM Model=UniDiffuser, Defense=DiffPure2026.0250.615613.2
COAVLM Model=LLaVA, Defense=Bit Reduction2026.0249.8676.723.2
COAVLM Model=MiniGPT-4, Defense=Bit Reduction2026.0249.4692.752.8
AdvDiffVLMVLM Model=UniDiffuser, Defense=ComDefend2026.024962.510.5
SGHA-Attack (Ours)VLM Model=LLaVA-NeXT, Defense=Bit Reduction2026.0248.4378.544.4
SGHA-Attack (Ours)VLM Model=LLaVA, Defense=ComDefend2026.0247.146.218
MF-iiVLM Model=MiniGPT-4, Defense=Bit Reduction2026.0245.4776.339.3
SGHA-Attack (Ours)VLM Model=LLaVA, Defense=DiffPure2026.0245.237.411.9
SGHA-Attack (Ours)VLM Model=MiniGPT-4, Defense=ComDefend2026.0244.9850.322.4
MF-iiVLM Model=LLaVA, Defense=JPEG Compression2026.0244.539.411.4
COAVLM Model=LLaVA, Defense=DiffPure2026.0244.3741.610.8
SGHA-Attack (Ours)VLM Model=MiniGPT-4, Defense=JPEG Compression2026.0243.2554.724.2
AdvDiffVLMVLM Model=LLaVA, Defense=JPEG Compression2026.0243.1334.46.8
AdvDiffVLMVLM Model=LLaVA, Defense=DiffPure2026.0242.4831.65.2
AdvDiffVLMVLM Model=LLaVA, Defense=Bit Reduction2026.0242.47346.6
COAVLM Model=MiniGPT-4, Defense=ComDefend2026.0242.1357.613
SGHA-Attack (Ours)VLM Model=MiniGPT-4, Defense=DiffPure2026.0242.0242.516.2
COAVLM Model=LLaVA, Defense=JPEG Compression2026.0241.3740.63.8
MF-iiVLM Model=LLaVA, Defense=DiffPure2026.0241.3118.12.6
COAVLM Model=LLaVA, Defense=ComDefend2026.0241.0435.12.6
MF-iiVLM Model=LLaVA, Defense=ComDefend2026.0240.6324.54.4
AdvDiffVLMVLM Model=LLaVA, Defense=ComDefend2026.0240.5426.12.1
COAVLM Model=MiniGPT-4, Defense=JPEG Compression2026.0239.9162.614.1
MF-iiVLM Model=MiniGPT-4, Defense=ComDefend2026.0239.731.17.3
SGHA-Attack (Ours)VLM Model=LLaVA-NeXT, Defense=JPEG Compression2026.0239.5449.215.8
COAVLM Model=MiniGPT-4, Defense=DiffPure2026.0239.148.89.5
AdvDiffVLMVLM Model=MiniGPT-4, Defense=Bit Reduction2026.0239.0144.715.8
AdvDiffVLMVLM Model=MiniGPT-4, Defense=DiffPure2026.0238.9938.510.8
AdvDiffVLMVLM Model=MiniGPT-4, Defense=ComDefend2026.0238.3331.46
AdvDiffVLMVLM Model=MiniGPT-4, Defense=JPEG Compression2026.0238.1338.89.8
MF-iiVLM Model=MiniGPT-4, Defense=JPEG Compression2026.0237.7635.58.9
MF-iiVLM Model=MiniGPT-4, Defense=DiffPure2026.0237.59256.4
COAVLM Model=LLaVA-NeXT, Defense=Bit Reduction2026.0237.0158.56.1
SGHA-Attack (Ours)VLM Model=LLaVA-NeXT, Defense=ComDefend2026.0236.637.57.4
MF-iiVLM Model=LLaVA-NeXT, Defense=JPEG Compression2026.0236.2226.53.5
SGHA-Attack (Ours)VLM Model=LLaVA-NeXT, Defense=DiffPure2026.0234.2325.63.8
AdvDiffVLMVLM Model=LLaVA-NeXT, Defense=ComDefend2026.0234.1727.70.4
AdvDiffVLMVLM Model=LLaVA-NeXT, Defense=Bit Reduction2026.0234.1327.41.3
COAVLM Model=LLaVA-NeXT, Defense=DiffPure2026.0233.7530.23.3
COAVLM Model=LLaVA-NeXT, Defense=ComDefend2026.0233.5831.91
COAVLM Model=LLaVA-NeXT, Defense=JPEG Compression2026.0233.2632.61.7
MF-iiVLM Model=LLaVA-NeXT, Defense=Bit Reduction2026.0233.0139.59.5
AdvDiffVLMVLM Model=LLaVA-NeXT, Defense=DiffPure2026.0232.9424.81.2
MF-iiVLM Model=LLaVA-NeXT, Defense=ComDefend2026.0232.7620.91.3
AdvDiffVLMVLM Model=LLaVA-NeXT, Defense=JPEG Compression2026.0232.6726.91.9
MF-iiVLM Model=LLaVA-NeXT, Defense=DiffPure2026.0231.7215.20.8