Research map / Architectures and efficiency / RNNs and normalization
Attribution-Guided Model Rectification of Unreliable Neural Network Behaviors
Peiyu Yang, Naveed Akhtar, Jiantong Jiang, Ajmal Mian
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 |
|---|---|---|---|
| use dynamic layer localization instead of final layer only (ablation) | rank-one rectifying (n=1) · CIFAR-10 Trojan | ASR (%): 2.57 -> 1.34 | Improved |
| use dynamic layer localization instead of final layer only (ablation) | rank-one rectifying (n=1) · ImageNet Trojan | ASR (%): 3.01 -> 1.61 | Improved |
| use dynamic layer localization instead of final layer only (ablation) | rank-one rectifying (n=1) · ImageNet Trojan | overall accuracy (%): 67.87 -> 66.77 | Hurt |
| use dynamic layer localization instead of final layer only (ablation) | rank-one rectifying (n=1) · CIFAR-10 spurious correlation | spurious-clean gap: 2.23 -> 1.86 | Improved |
| use dynamic layer localization instead of final layer only (ablation) | rank-one rectifying (n=1) · BlockMNIST | feature leakage: 2.655 -> 2.417 | Improved |
| use dynamic layer localization instead of final layer only (ablation) | rank-one rectifying · ISIC | spurious-clean gap: 4.50 -> 1.50 | Improved |
| rank-one rectifying instead of fine-tuning | Trojaned CIFAR-10 model · CIFAR-10 Trojan | ASR (%): 13.22 -> 0.26 | Improved |
| rank-one rectifying instead of P-ClArC | Trojaned CIFAR-10 model · CIFAR-10 Trojan | overall accuracy (%): 89.97 -> 93.61 | Improved |
| increase cleansed samples from 1 to 20 (ablation) | Dyn. rectifying · CIFAR-10 Trojan | ASR (%): 1.34 -> 0.26 | Improved |
| increase cleansed samples from 1 to 20 (ablation) | Dyn. rectifying · ImageNet Trojan | overall accuracy (%): 66.77 -> 68.84 | Improved |
| fine-tune instead of rectify | benign model · CIFAR-10 spurious correlation | spurious-clean gap: 5.58 -> 10.65 | Hurt |
| add IG regularization | benign BlockMNIST model · BlockMNIST | accuracy: 99.17 -> 94.14 | Hurt |
Closest papers by meaning
- ModHiFi: Identifying High Fidelity predictive components for Model Modification (NeurIPS 2025 · similarity 0.86)
- GNNs Also Deserve Editing, and They Need It More Than Once (ICML 2024 · similarity 0.85)
- DisCEdit: Model Editing by Identifying Discriminative Components (NeurIPS 2024 · similarity 0.85)
- Learning Where to Edit Vision Transformers (NeurIPS 2024 · similarity 0.85)
- Model Diagnosis and Correction via Linguistic and Implicit Attribute Editing (CVPR 2025 · similarity 0.85)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.84)
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.
