Research map / Adversarial robustness and security / Black-box and transfer attacks
Defense Against Adversarial Attacks Using High-Level Representation Guided Denoiser
Read the paperCheck 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 |
|---|---|---|---|
| replace pixel-level reconstruction loss with logits-guided loss (ablation) | DUNET denoiser on Inception V3 · ImageNet WhiteTestSet ε=16 | accuracy: 13.8% -> 69.2% | Helped |
| replace pixel-level reconstruction loss with logits-guided loss (ablation) | DUNET denoiser on Inception V3 · ImageNet BlackTestSet ε=16 | accuracy: 55.7% -> 72.2% | Helped |
| guide denoising with logits rather than top convolutional features (ablation) | HGD on Inception V3 · ImageNet WhiteTestSet ε=16 | accuracy: 67.4% -> 69.2% | Helped |
| guide denoising with logits rather than class-label loss (ablation) | HGD on Inception V3 · ImageNet WhiteTestSet ε=16 | accuracy: 73.2% -> 69.2% | Hurt |
| guide denoising with logits rather than class-label loss (ablation) | HGD on Inception V3 · ImageNet BlackTestSet ε=16 | accuracy: 71.1% -> 72.2% | Helped |
| add U-net lateral connections and residual noise prediction instead of a bottleneck autoencoder (ablation) | image denoiser on Inception V3 · ImageNet clean | accuracy: 58.3% -> 75.3% | Helped |
| use logits-guided denoising instead of no defense | Inception V3 · ImageNet WhiteTestSet ε=16 | accuracy: 14.4% -> 69.2% | Helped |
| use logits-guided denoising instead of no defense | Inception V3 · ImageNet BlackTestSet ε=16 | accuracy: 41.0% -> 72.2% | Helped |
| use IncV3-guided LGD on Resnet instead of no defense | Resnet · ImageNet WhiteTestSet ε=16 | accuracy: 38.4% -> 71.7% | Helped |
| train LGD without examples from 250 ImageNet classes | LGD · ImageNet unseen-class test | accuracy: 15.3% -> 65.7% | Helped |
| use HGD rather than ensemble adversarial training | Inception V3 defense · ImageNet WhiteTestSet ε=16 | accuracy: 58.0% -> 69.2% | Helped |
| use HGD rather than ensemble adversarial training | Inception V3 defense · ImageNet BlackTestSet ε=16 | accuracy: 62.0% -> 72.2% | Helped |
Closest papers by meaning
- Robust Overfitting Does Matter: Test-Time Adversarial Purification With FGSM (CVPR 2024 · similarity 0.86)
- PubDef: Defending Against Transfer Attacks From Public Models (ICLR 2024 · similarity 0.85)
- First Line of Defense: A Robust First Layer Mitigates Adversarial Attacks (AAAI 2025 · similarity 0.85)
- AdvAD: Exploring Non-Parametric Diffusion for Imperceptible Adversarial Attacks (NeurIPS 2024 · similarity 0.85)
- Adversarial Training Should Be Cast as a Non-Zero-Sum Game (ICLR 2024 · similarity 0.85)
- One Stone, Two Birds: Enhancing Adversarial Defense Through the Lens of Distributional Discrepancy (ICML 2025 · similarity 0.85)
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.