Image Quality Assessment on GenAI-Bench
98.588VQAScoreMGSIA
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MGSIAJudge Backbone=LLaVA-1.5-7b, Attack Category=Ours, Perturbation Budget=8/2552026.06 | 98.588 | 98.062 | 93.999 | 89.927 | 88.347 | 88.253 | |
| MGSIAJudge Backbone=Qwen2.5-VL-7b, Attack Category=Ours, Perturbation Budget=8/2552026.06 | 89.042 | 96.639 | 96.579 | 93.227 | 94.567 | 93.983 | |
| Target Text EmbeddingJudge Backbone=Qwen2.5-VL-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 86.694 | 84.408 | 93.814 | 84.76 | 89.56 | 86.173 | |
| Target Text EmbeddingJudge Backbone=LLaVA-1.5-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 79.823 | 91.997 | 91.016 | 89.16 | 85.493 | 84.961 | |
| Target Text OverlayJudge Backbone=Qwen2.5-VL-7b, Attack Category=Handcrafted Perturbation, Perturbation Budget=8/2552026.06 | 75.975 | 83.32 | 88.32 | 82.48 | 83.887 | 81.413 | |
| AttackVLM-itJudge Backbone=Qwen2.5-VL-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 75.93 | 82.977 | 88.265 | 82.507 | 83.547 | 81.78 | |
| Chain of AttackJudge Backbone=Qwen2.5-VL-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 75.702 | 81.343 | 87.923 | 81.667 | 82.487 | 80.393 | |
| Gaussian NoiseJudge Backbone=Qwen2.5-VL-7b, Attack Category=Image Degradation, Perturbation Budget=8/2552026.06 | 75.558 | 82.27 | 87.805 | 82.673 | 83.12 | 81.8 | |
| AttackVLM-iiJudge Backbone=Qwen2.5-VL-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 75.544 | 82.125 | 87.665 | 82.6 | 83.553 | 81.927 | |
| PaddingJudge Backbone=Qwen2.5-VL-7b, Attack Category=Handcrafted Perturbation, Perturbation Budget=8/2552026.06 | 75.52 | 81.813 | 87.657 | 82.06 | 83.02 | 81.2 | |
| BenignJudge Backbone=Qwen2.5-VL-7b2026.06 | 75.4 | 82.383 | 87.839 | 82.033 | 83.36 | 81 | |
| Bounding BoxJudge Backbone=Qwen2.5-VL-7b, Attack Category=Handcrafted Perturbation, Perturbation Budget=8/2552026.06 | 74.817 | 81.659 | 87.833 | 82.04 | 83.047 | 80.607 | |
| Target Text OverlayJudge Backbone=LLaVA-1.5-7b, Attack Category=Handcrafted Perturbation, Perturbation Budget=8/2552026.06 | 70.898 | 88.289 | 85.836 | 89.313 | 84.32 | 85.68 | |
| Gaussian BlurJudge Backbone=Qwen2.5-VL-7b, Attack Category=Image Degradation, Perturbation Budget=8/2552026.06 | 67.354 | 78.856 | 86.569 | 74.793 | 73.66 | 68.987 | |
| Bounding BoxJudge Backbone=LLaVA-1.5-7b, Attack Category=Handcrafted Perturbation, Perturbation Budget=8/2552026.06 | 64.612 | 87.391 | 84.458 | 89.173 | 84.12 | 84.527 | |
| Chain of AttackJudge Backbone=LLaVA-1.5-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 64.12 | 87.684 | 84.899 | 89.147 | 84 | 84.3 | |
| PaddingJudge Backbone=LLaVA-1.5-7b, Attack Category=Handcrafted Perturbation, Perturbation Budget=8/2552026.06 | 63.594 | 87.244 | 84.843 | 89.04 | 84.147 | 84.607 | |
| BenignJudge Backbone=LLaVA-1.5-7b2026.06 | 63.001 | 87.405 | 84.827 | 89.2 | 84.347 | 84.76 | |
| AttackVLM-itJudge Backbone=LLaVA-1.5-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 62.789 | 85.772 | 83.745 | 88.96 | 83.16 | 83.567 | |
| Gaussian NoiseJudge Backbone=LLaVA-1.5-7b, Attack Category=Image Degradation, Perturbation Budget=8/2552026.06 | 62.457 | 87.131 | 84.742 | 89.213 | 84.107 | 84.747 | |
| Gaussian BlurJudge Backbone=LLaVA-1.5-7b, Attack Category=Image Degradation, Perturbation Budget=8/2552026.06 | 62.052 | 85.474 | 83.607 | 89.267 | 83.573 | 84.593 | |
| AttackVLM-iiJudge Backbone=LLaVA-1.5-7b, Attack Category=Learnable Perturbation, Perturbation Budget=8/2552026.06 | 61.872 | 86.925 | 83.694 | 88.853 | 83.427 | 83.607 |