Research map / Adversarial robustness and security / Black-box and transfer attacks
On Generating Transferable Targeted Perturbations
Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Fatih Porikli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add augmentations during generator training (ablation) | TTP · Per-target average, VGG16 (SIN) target, ResNet50 source | target top-1 accuracy (%): 17.8 -> 30.3 | Improved |
| add augmentations during generator training (ablation) | TTP · Per-target average, VGG16 (natural) target, ResNet50 source | target top-1 accuracy (%): 63.5 -> 78.8 | Improved |
| use a ResNet ensemble as source instead of ResNet50 (ablation) | TTP · Per-target average, VGG16 (SIN) target | target top-1 accuracy (%): 30.3 -> 53.6 | Improved |
| use a ResNet ensemble as source instead of ResNet50 (ablation) | TTP · Per-target average, VGG16 (natural) target | target top-1 accuracy (%): 78.8 -> 89.1 | Improved |
| train with distribution matching loss only instead of cross-entropy (GAP) or relativistic cross-entropy (CDA) (ablation) | generator attack · ResNet50 to VGG16 (SIN) | target transferability: more than 100% | Improved |
| add Gaussian smoothing operator W (3x3) (ablation) | TTP · ResNet50 to VGG16 (SIN) | target transferability | Improved |
| add augmentations with both distribution matching losses (ablation) | TTP · ResNet50 to VGG16 (SIN) | target transferability | Improved |
| add neighbor similarity loss (ablation) | TTP · ResNet50 to VGG16 (SIN) | target transferability | Improved |
| use TTP instead of CDA | generative attack · 10-targets all-source, ResNet50 to VGG19_BN | target top-1 accuracy (%): 64.58 -> 78.15 | Improved |
| use TTP instead of CDA against stylized-trained ResNet50 (100 targets) | generative attack · 100-targets, SIN target | target top-1 accuracy (%): 7.6 -> 31.0 | Improved |
| raise perturbation budget from 16 to 32 | TTP · 10-targets, SIN target | target top-1 accuracy (%): 30.17 -> 41.52 | Improved |
| apply median blur defense on clean images (VGG19_BN) | VGG19_BN · ImageNet val clean accuracy | top-1 accuracy (%): 74.24 -> 53.86 | Hurt |
Closest papers by meaning
- Boosting Transferability of Targeted Adversarial Examples via Hierarchical Generative Networks (ECCV 2022 · similarity 0.93)
- AIM: Additional Image Guided Generation of Transferable Adversarial Attacks (AAAI 2025 · similarity 0.92)
- Towards Transferable Adversarial Attacks with Centralized Perturbation (AAAI 2024 · similarity 0.91)
- Improving Transferable Targeted Adversarial Attacks with Model Self-Enhancement (CVPR 2024 · similarity 0.89)
- Boosting Adversarial Transferability through Augmentation in Hypothesis Space (CVPR 2025 · similarity 0.89)
- Everywhere Attack: Attacking Locally and Globally to Boost Targeted Transferability (AAAI 2025 · similarity 0.89)
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.
