Research map / Adversarial robustness and security / Attacks on text and graphs
Understanding Adversarial Examples From the Mutual Influence of Images and Perturbations
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use COCO proxy images instead of original ImageNet training data to train targeted UAPs | Proposed targeted UAP · ImageNet test, 8 target classes | targeted fooling ratio (%): ImageNet: 48.6 ± 13.3; COCO: 47.2 ± 13.1 (AlexNet); ImageNet: 75.0 ± 7.8; COCO: 75.1 ± 12.3 (VGG16) | No clear effect |
| Use COCO proxy images to generate non-targeted UAPs | Proposed UAP · ImageNet test | non-targeted fooling ratio (%): 89.9 vs 89.04 (AlexNet); 76.8 vs 75.28 (GoogleNet); 92.2 vs 71.59 (VGG16); 91.6 vs 72.84 (VGG19); 79.9 vs 60.72 (ResNet1 | Helped |
| Use L^t_CL2 instead of cross-entropy loss (ablation) | Proposed targeted UAP · ImageNet, target class sea lion | targeted fooling ratio (%): AlexNet: 55.4 ± 1.0 -> 49.4 ± 1.2; GoogleNet: 55.2 ± 2.2 -> 58.4 ± 2.2; VGG16: 75.9 ± 0.9 -> 82.8 ± 0.7; VGG19: 70.8 ± 1 | Mixed |
| Use L^t_CL1 instead of cross-entropy loss (ablation) | Proposed targeted UAP · ImageNet, target class sea lion | targeted fooling ratio (%): 55.4 ± 1.0 -> 57.6 ± 1.4 (AlexNet); 55.2 ± 2.2 -> 57.9 ± 2.3 (GoogleNet); 75.9 ± 0.9 -> 80.3 ± 0.5 (VGG16); 70.8 ± 1.1 - | Mixed |
| Use L_L instead of cross-entropy loss (ablation) | Proposed targeted UAP · ImageNet, target class sea lion | targeted fooling ratio (%): 55.4 ± 1.0 -> 47.1 ± 1.1 (AlexNet); 55.2 ± 2.2 -> 56.9 ± 1.1 (GoogleNet); 75.9 ± 0.9 -> 79.0 ± 0.6 (VGG16); 70.8 ± 1.1 - | Helped |
| Use L^t_CL2 instead of L^t_CL1 (ablation) | Proposed targeted UAP · ImageNet, target class sea lion | targeted fooling ratio (%): 57.6 ± 1.4 -> 49.4 ± 1.2 (AlexNet); 57.9 ± 2.3 -> 58.4 ± 2.2 (GoogleNet); 80.3 ± 0.5 -> 82.8 ± 0.7 (VGG16); 75.5 ± 0.6 - | Mixed |
| Use white images or Gaussian noise instead of proxy images | Proposed targeted UAP · Targeted UAP generation | targeted fooling ratio | Hurt |
| Use a random source dataset as proxy data | Targeted universal perturbation · ImageNet test | targeted fooling ratio: 53.4% vs 52% | Helped |
Closest papers by meaning
- Art of Singular Vectors and Universal Adversarial Perturbations (CVPR 2018 · similarity 0.91)
- Universal Adversarial Training (AAAI 2020 · similarity 0.90)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.89)
- Theoretical Understanding of Learning from Adversarial Perturbations (ICLR 2024 · similarity 0.89)
- Generate Universal Adversarial Perturbations for Few-Shot Learning (NeurIPS 2024 · similarity 0.87)
- Improving Generalization of Universal Adversarial Perturbation via Dynamic Maximin Optimization (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.