Research map / Adversarial robustness and security / Black-box and transfer attacks
Any Target Can be Offense: Adversarial Example Generation via Generalized Latent Infection
Youheng Sun, Shengming Yuan, Xuanhan Wang, Lianli Gao, Jingkuan Song
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 |
|---|---|---|---|
| condition on target images with latent infection instead of class labels, enabling unknown targets | GAKer vs HGN · Unknown classes, Res-50 -> VGG-19 | targeted attack success rate (%): 0.15 -> 26.02 | Improved |
| select known classes with similarity greedy algorithm (most dissimilar features) instead of random (ablation) | GAKer · Res-50 -> Res-152 | targeted attack success rate (%): +16.52% | Improved |
| select training images with lowest classification loss instead of random or highest loss (ablation) | GAKer · unknown classes, Res-50 | targeted attack success rate (%) | Improved |
| increase number of known classes to 500 (ablation) | GAKer · unknown classes, Res-50 | targeted attack success rate (%): 49.45 white-box on 500 unknown classes at N=500 | No clear effect |
| use very few known classes (N=10 or 50) (ablation) | GAKer · unknown classes | targeted attack success rate (%) | Hurt |
| use more than 325 images per known class (ablation) | GAKer · unknown classes, Res-50 | targeted attack success rate (%): peak 41.69 at M=325 | Hurt |
| use fewer than 325 images per known class (ablation) | GAKer · unknown classes, Res-50 | targeted attack success rate (%) | Hurt |
| GAKer vs ESMA on known classes with Res-50 surrogate | GAKer · known classes, Res-50 surrogate | targeted attack success rate (%) avg: 61.37 -> 58.71 | Hurt |
| GAKer-10 vs single-target TTP-10 with VGG-19 surrogate | GAKer · known classes (10), VGG-19 surrogate | targeted attack success rate (%) avg: 36.98 -> 25.51 | Hurt |
| train one generator for all target classes instead of retraining per class batch | GAKer vs TTP · training time | relative training time: 100% -> 1% | Improved |
Closest papers by meaning
- GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification (ICLR 2020 · similarity 0.89)
- Constructing Unrestricted Adversarial Examples with Generative Models (NeurIPS 2018 · similarity 0.89)
- GAMA: Generative Adversarial Multi-Object Scene Attacks (NeurIPS 2022 · similarity 0.88)
- Dynamic Generative Targeted Attacks With Pattern Injection (CVPR 2023 · similarity 0.88)
- Adversarial Example Games (NeurIPS 2020 · similarity 0.88)
- Towards Imperceptible and Robust Adversarial Example Attacks Against Neural Networks (AAAI 2018 · similarity 0.88)
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.
