Research map / Adversarial robustness and security / Attacks on text and graphs
Downstream-agnostic Adversarial Examples
Ziqi Zhou, Shengshan Hu, Ruizhi Zhao, Qian Wang, Leo Yu Zhang, Junhui Hou, Hai Jin
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 |
|---|---|---|---|
| use Adv-PER instead of UAP/UPGD/FFF/SSP/NAG/PAP baselines | universal perturbation attack on pre-trained encoders · STL10 downstream, ImageNet encoders | ASR (%): SimCLR 55.87 vs PAP-ugs 53.56, UAP 48.41 | Improved |
| use Adv-PAT instead of UA-PAT patch | universal patch attack · STL10 downstream | ASR (%): BYOL 34.62 -> 88.60 | Improved |
| use ImageNet instead of CIFAR10 as the attacker's surrogate dataset (ablation) | AdvEncoder · Adv-PER/Adv-PAT, 14 encoders | ASR (%) | Improved |
| increase number of surrogate samples (ablation) | Adv-PER · ImageNet encoders SimCLR and MoCo v2+, CIFAR10 downstream | ASR (%) | Improved |
| increase surrogate data size (ablation) | Adv-PAT · ImageNet encoders, CIFAR10 downstream | ASR (%) | No clear effect |
| add high-frequency component (HFC) loss and generator (ablation) | Adv-PER and Adv-PAT · STL10 downstream | ASR (%) | Improved |
| apply Gaussian noise corruption defense (ablation) | AdvEncoder · CIFAR10 | attack accuracy: noise 0.03 | No clear effect |
| full fine-tuning of MoCo v2+ encoder and pruning rate 0.1 to 0.7 as defense (ablation) | AdvEncoder · CIFAR10, STL10, GTSRB, ImageNet | ASR | No clear effect |
| adversarial training of the encoder as defense (ablation) | AdvEncoder · ImageNet-trained SimCLR, CIFAR10 downstream | ASR | Mixed |
Closest papers by meaning
- AdvDiffuser: Natural Adversarial Example Synthesis with Diffusion Models (ICCV 2023 · similarity 0.85)
- Adversarial Machine Learning at Scale (ICLR 2017 · similarity 0.85)
- Generalizable Data-Free Objective for Crafting Universal Adversarial Perturbations (TPAMI 2018 · similarity 0.85)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.85)
- Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models (AAAI 2025 · similarity 0.84)
- RAEncoder: A Label-Free Reversible Adversarial Examples Encoder for Dataset Intellectual Property Protection (CVPR 2025 · similarity 0.84)
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.
