Research map / Adversarial robustness and security / Attacks on text and graphs
Deflecting Adversarial Attacks With Pixel Deflection
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 |
|---|---|---|---|
| apply pixel deflection and wavelet denoising | ResNet-50 · ImageNet val, FGSM | Top-1 accuracy: 20.0 -> 81.5 | Helped |
| apply pixel deflection and wavelet denoising | ResNet-50 · ImageNet val, C&W | Top-1 accuracy: 04.8 -> 98.0 | Helped |
| add wavelet denoising to pixel deflection (ablation) | Pixel Deflection · ImageNet val, FGSM | Top-1 accuracy: 79.7 -> 81.5 | Helped |
| add JPEG compression to pixel deflection and wavelet denoising (ablation) | Pixel Deflection + Wavelet Denoising · ImageNet val, FGSM | Top-1 accuracy: 81.5 -> 78.8 | Hurt |
| replace BayesShrink with hard thresholding (ablation) | Pixel Deflection + Wavelet Denoising · ImageNet val, FGSM | Top-1 accuracy: 81.5 -> 35.9 | Hurt |
| replace BayesShrink with VisuShrink (ablation) | Pixel Deflection + Wavelet Denoising · ImageNet val, FGSM | Top-1 accuracy: 81.5 -> 63.8 | Hurt |
| replace BayesShrink with SUREShrink (ablation) | Pixel Deflection + Wavelet Denoising · ImageNet val, FGSM | Top-1 accuracy: 81.5 -> 79.7 | Hurt |
| increase pixel deflections from 100 to 10K (ablation) | Pixel Deflection + Wavelet Denoising · ImageNet val, clean | Top-1 accuracy: 98.1 -> 80.3 | Hurt |
| increase pixel deflections from 100 to 10K (ablation) | Pixel Deflection + Wavelet Denoising · ImageNet val, JSMA | Top-1 accuracy: 93.0 -> 67.7 | Hurt |
| increase window size from 10 to 100 (ablation) | Pixel Deflection + Wavelet Denoising · ImageNet val, clean | Top-1 accuracy: 98.1 -> 94.4 | Hurt |
| replace uniform pixel sampling with Gaussian sampling (ablation) | Pixel Deflection · ImageNet val, six attacks (average; window 10) | Top-1 accuracy: 87.5 -> 81.4 | Hurt |
| replace random neighborhood pixel values with neighborhood mean values (ablation) | Pixel Deflection · ImageNet val, six attacks (average; window 10) | Top-1 accuracy: 87.5 -> 72.3 | Hurt |
Closest papers by meaning
- PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples (ICLR 2018 · similarity 0.90)
- Feature Denoising for Improving Adversarial Robustness (CVPR 2019 · similarity 0.89)
- Towards Achieving Adversarial Robustness by Enforcing Feature Consistency Across Bit Planes (CVPR 2020 · similarity 0.88)
- Improving Accuracy-robustness Trade-off via Pixel Reweighted Adversarial Training (ICML 2024 · similarity 0.87)
- IRAD: Implicit Representation-driven Image Resampling against Adversarial Attacks (ICLR 2024 · similarity 0.86)
- Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature (AAAI 2025 · similarity 0.86)
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.