Research map / Adversarial robustness and security / Attacks on text and graphs
Attacks Meet Interpretability: Attribute-steered Detection of Adversarial Samples
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 |
|---|---|---|---|
| Use AmI instead of feature squeezing | VGG-Face adversarial detection · Targeted attacks and FGSM/BIM | detection accuracy: AmI achieves more than 90% accuracy for most attacks; FS best result 77% | Helped |
| Use AmI instead of feature squeezing | VGG-Face · VF benign inputs | false positive rate: 23.32% -> 9.91% | Helped |
| Use AmI instead of feature squeezing | VGG-Face · CelebA benign inputs | false positive rate: 28.43% -> 8.97% | Helped |
| Use attribute witness activations rather than fc7 face-descriptor activations | VGG-Face attribute detection · VF, Left Eye | accuracy: 0.830 -> 0.940 | Helped |
| Use attribute substitution alone instead of bidirectional witness extraction (ablation) | AmI · VF benign inputs | false positive rate: 9.91% -> 20.41% | Hurt |
| Use attribute preservation alone instead of bidirectional witness extraction (ablation) | AmI · VF benign inputs | false positive rate: 9.91% -> 30.61% | Hurt |
| Use non-witness weakening alone, without witness strengthening (ablation) | AmI · VF benign inputs | false positive rate: 9.91% -> 7.87% | Helped |
| Use non-witness weakening alone, without witness strengthening (ablation) | AmI · FGSM | detection accuracy: 0.91 -> 0.86 | No clear effect |
| Use witness strengthening alone, without non-witness weakening (ablation) | AmI · FGSM | detection accuracy: 0.91 -> 0.46 | Hurt |
| Use witness strengthening alone, without non-witness weakening (ablation) | AmI · VF benign inputs | false positive rate: 9.91% -> 2.33% | Helped |
| Exclude left-eye witnesses (ablation) | AmI · Glasses (First) | detection accuracy: 0.85 -> 0.75 | Hurt |
| Exclude right-eye witnesses (ablation) | AmI · Glasses (First) | detection accuracy: 0.85 -> 0.73 | Hurt |
Closest papers by meaning
- Robust Feature-Level Adversaries are Interpretability Tools (NeurIPS 2022 · similarity 0.89)
- Adv-Attribute: Inconspicuous and Transferable Adversarial Attack on Face Recognition (NeurIPS 2022 · similarity 0.88)
- Proper Network Interpretability Helps Adversarial Robustness in Classification (ICML 2020 · similarity 0.88)
- On Detecting Adversarial Perturbations (ICLR 2017 · similarity 0.87)
- Theoretical Understanding of Learning from Adversarial Perturbations (ICLR 2024 · similarity 0.86)
- Detecting Adversarial Data Using Perturbation Forgery (CVPR 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.