Research map / Adversarial robustness and security / Attacks on text and graphs
Fooling Network Interpretation in Image Classification
Akshayvarun Subramanya, Vipin Pillai, Hamed Pirsiavash
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 Grad-CAM suppression term to targeted patch loss (ablation) | Adversarial patch (VGG19-BN) · ImageNet val | energy ratio: 76.26 -> 6.80 | Improved |
| add Grad-CAM suppression term to targeted patch loss (ablation) | Adversarial patch (VGG19-BN) · ImageNet val | target accuracy: 99.98 -> 77.88 | Hurt |
| add Grad-CAM suppression term to non-targeted patch loss (ablation) | Adversarial patch (VGG19-BN) · ImageNet val | energy ratio: 50.87 -> 2.61 | Improved |
| add suppression term, ResNet-34 patch at top-right (ablation) | Adversarial patch · ImageNet val 5,000 images | energy ratio: 61.9 -> 8.2 | Improved |
| add suppression term, DenseNet-121 patch at top-right (ablation) | Adversarial patch · ImageNet val 5,000 images | energy ratio: 71.3 -> 5.3 | Improved |
| encourage Grad-CAM at the opposite corner (uniform heatmap) (ablation) | Adversarial patch · ImageNet val | top-left energy: 76.96 -> 14.99 | Improved |
| add suppression term on GAIN_ext model (ablation) | Adversarial patch · PASCAL VOC-2012 | energy ratio: 37.90 -> 3.2 | Improved |
| evaluate Grad-CAM-trained patch with Occluding Patch (ablation) | Adversarial patch · PASCAL VOC-2012 GAIN_ext | energy: 80.44 -> 31.59 | Improved |
| add Grad-CAM-fooling term to regular PGD examples (ablation) | PGD adversarial examples · ImageNet val | localization error: 67.74 -> 76.02 | Improved |
| add Grad-CAM-fooling term to regular PGD examples (ablation) | PGD adversarial examples · ImageNet val | histogram intersection: 0.77 -> 0.64 | Improved |
| add suppression term, universal patch for class cow (ablation) | Universal patch · PASCAL VOC-2012 GAIN_ext | target accuracy: 100 -> 34.3 | Hurt |
Closest papers by meaning
- Fooling Neural Network Interpretations via Adversarial Model Manipulation (NeurIPS 2019 · similarity 0.90)
- Proper Network Interpretability Helps Adversarial Robustness in Classification (ICML 2020 · similarity 0.89)
- Attack to Fool and Explain Deep Networks (TPAMI 2021 · similarity 0.88)
- What Does It Mean to Learn in Deep Networks? And, How Does One Detect Adversarial Attacks? (CVPR 2019 · similarity 0.87)
- Learning to Attack: Adversarial Transformation Networks (AAAI 2018 · similarity 0.87)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (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.
