Research map / Adversarial robustness and security / Certified and empirical robustness
Natural Jammers: When Non-Robust Features Become Antagonistic
Jiaming Zhang, Fuyao Zhang, Pengjun Xie, Wei Yang Bryan Lim
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that non-robust features can be antagonistic, acting as "Natural Jammers" that mislead standard models on clean, unmodified images, rather than only serving as helpful shortcuts.
Problem
Prior work frames non-robust features as opportunistic shortcuts that help standard models on clean data; the paper asks whether such features can instead actively mislead standard models on clean images.
Approach
Reframes non-robust features as potentially antagonistic: identifies "Natural Jammers," clean unmodified images whose non-robust cues push standard models toward wrong predictions, using disagreement between a standard model and a robust model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove high-frequency components with a low-pass filter (radius 50) (ablation) | standard ResNet-50 on the Jammed Set · ImageNet val | accuracy: 0.0% -> 31.8% | Improved |
| remove high-frequency components with a low-pass filter (radius 30) (ablation) | standard ResNet-50 on the Spurious Set · ImageNet val | accuracy: 100.0% -> 23.7% | Hurt |
| remove high-frequency components with a low-pass filter (radius 50) (ablation) | standard ResNet-50 on the Common Set · ImageNet val | accuracy: 100.0% -> 85.1% | Hurt |
| zoom in with center crop (ratio 1.0 to 0.5) (ablation) | standard ResNet-50 on the Jammed Set · ImageNet val | accuracy | Improved |
| zoom in with center crop (ratio 1.0 to 0.5) (ablation) | standard ResNet-50 on the Spurious Set · ImageNet val | accuracy | Hurt |
| zoom in with center crop (ratio 1.0 to 0.5) (ablation) | standard ResNet-50 on the Common Set · ImageNet val | accuracy: 100% -> 79.9% | Hurt |
| graft jammer high frequencies onto a Common-set host image (ablation) | standard ResNet-50 · ImageNet val | transferred toxicity: 0.17% transferred vs 99.9% on own host | No clear effect |
| apply 2% edge crop and resize (ablation) | Jammed Set of standard ResNet-50 · ImageNet val | set size |J_T|: 2,140 -> 2,289 | No clear effect |
| apply 2% edge crop and resize (ablation) | Jammed Set of standard ResNet-50 · ImageNet val | IoU with original Jammed Set: 1.00 -> 0.46 | Hurt |
| increase crop from 2% to 8% (ablation) | Jammed Set of standard ResNet-50 · ImageNet val | IoU with original Jammed Set: 0.46 -> 0.35 | Hurt |
| use stronger adversarial robustness budget for the robust lens (l2 eps 0.5 to 3.0) (ablation) | ResNet-18 quad-set partition · ImageNet val | Jammed set share: 5.28% -> 3.85% | Hurt |
| change the architecture from ResNet-18 to WRN-50-2 or ViT (ablation) | Jammed Set prevalence · ImageNet val | Jammed set share: 4.90% -> 4.33% (RN-18 vs WRN-50-2, l2 eps=1) | No clear effect |
Surprising findings
- Non-robust features can actively mislead standard models on clean, unmodified images, not only act as helpful shortcuts.
Closest papers by meaning
- Finding Biological Plausibility for Adversarially Robust Features via Metameric Tasks (ICLR 2022 · similarity 0.85)
- Which Models have Perceptually-Aligned Gradients? An Explanation via Off-Manifold Robustness (NeurIPS 2023 · similarity 0.85)
- Revisiting Visual Model Robustness: A Frequency Long-Tailed Distribution View (NeurIPS 2023 · similarity 0.85)
- Suppress and Diversify: Refining Robust Pathways for Corruption Robustness (ICML 2026 · similarity 0.85)
- Detecting Brittle Decisions for Free: Leveraging Margin Consistency in Deep Robust Classifiers (NeurIPS 2024 · similarity 0.83)
- A Sober Look at the Robustness of CLIPs to Spurious Features (NeurIPS 2024 · similarity 0.83)
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.
