Research map / Adversarial robustness and security / Attacks on text and graphs
On Detecting Adversarial Perturbations
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 |
|---|---|---|---|
| Add a detector subnetwork trained on static adversarial examples | ResNet-32 · CIFAR10, static adversaries that reduce classifier accuracy below 30% | detector accuracy: above 80\% | Helped |
| Add a detector subnetwork trained on static adversarial examples | ResNet-32 · CIFAR10, static adversaries that reduce classifier accuracy below 10% | detector accuracy: above 90\% | Helped |
| Train the detector with dynamic adversaries rather than only static adversaries (ablation) | CIFAR10 detector · CIFAR10, dynamic adversary (ε=1; σ from 0.0 to 1.0) | detector accuracy: more than 70\% | Helped |
| Use a static detector instead of a detector trained against dynamic adversaries (ablation) | CIFAR10 detector · CIFAR10, dynamic adversary at σ=0.3 and σ=0.4 | detector accuracy | Hurt |
| Attach the detector at AD(2) rather than other positions (ablation) | CIFAR10 detector for Fast and Iterative adversaries · CIFAR10, Fast and Iterative adversaries | detector accuracy | Helped |
| Attach the detector at AD(4) rather than other positions (ablation) | CIFAR10 detector for DeepFool adversaries · CIFAR10, DeepFool adversaries | detector accuracy | Helped |
| Train the detector with a smaller ε rather than a larger ε (ablation) | CIFAR10 detector · CIFAR10, same adversary evaluated at different ε | detector accuracy | Mixed |
| Train the detector on Iterative rather than Fast adversaries (ablation) | CIFAR10 detector · CIFAR10, transfer between Iterative and Fast adversaries | detector accuracy | Helped |
| Train the detector on DeepFool rather than Iterative adversaries (ablation) | CIFAR10 detector · CIFAR10, transfer to other adversaries | detector accuracy | Hurt |
| Train the detector on DeepFool adversaries (ablation) | 10-class ImageNet detector · 10-class ImageNet, transfer to other adversaries | detector accuracy | Helped |
| Train the detector with ε=1200 rather than ε=400 (ablation) | 10-class ImageNet detector · 10-class ImageNet, Iterative ℓ2 adversary at ε=400 | detector accuracy: 78\% | Helped |
| Train the detector for Iterative ℓ2 at ε=400 rather than ε=1200 (ablation) | 10-class ImageNet detector · 10-class ImageNet, Iterative ℓ2 adversary at ε=400 | detector accuracy | Hurt |
Closest papers by meaning
- Defense Against Universal Adversarial Perturbations (CVPR 2018 · similarity 0.91)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.90)
- Universal Adversarial Perturbations (CVPR 2017 · similarity 0.90)
- Detecting Adversarial Data Using Perturbation Forgery (CVPR 2025 · similarity 0.89)
- Theoretical Understanding of Learning from Adversarial Perturbations (ICLR 2024 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
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.