Research map / Adversarial robustness and security / Certified and empirical robustness
FastJSMA: Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling
Zhenghao Gao, Shengjie Xu, Zijing Li, Meixi Chen, Chaojian Yu, Yuanjie Shao, Changxin Gao
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 class excitation gradient g+ to the suppression-only attack (ablation) | FastJSMA · CIFAR-10 | ASR (%): 35.75 -> 43.98 | Improved |
| add class suppression gradient g- to the excitation-only attack (ablation) | FastJSMA · CIFAR-10 | ASR (%): 30.66 -> 43.98 | Improved |
| use suppression gradient alone instead of excitation gradient alone (ablation) | FastJSMA · CIFAR-10 | ASR (%): 30.66 -> 35.75 | Improved |
| use both gradients instead of a single one: effect on generation time (ablation) | FastJSMA · CIFAR-10 | time (µs): 339 vs 432 and 443 | No clear effect |
| replace full-Jacobian JSMANT with FastJSMA gradient decoupling | JSMANT · CIFAR-10 | time (µs): 11,769 -> 339 | Improved |
| replace full-Jacobian JSMANT with FastJSMA gradient decoupling | JSMANT · CIFAR-100 | time (µs): 61,166 -> 751 | Improved |
| replace full-Jacobian JSMANT with FastJSMA gradient decoupling | JSMANT · CIFAR-100 | ASR (%): 43.87 -> 45.30 | Improved |
| use FastJSMA instead of weighted WJSMA | WJSMA · CIFAR-10 | ASR (%): 51.56 -> 43.98 | Hurt |
| increase max attempt (class probing) from 1 to 10 (ablation) | FastJSMA · CIFAR-10 | ASR (%): 35.2% at attempt=1 to 44.7% at attempt=10 | Improved |
| increase max attempt from 1 to 10: effect on time (ablation) | FastJSMA · CIFAR-10 | time (µs): 314 -> 417 | Hurt |
| increase number of perturbed pixels from 2 to 50 (ablation) | FastJSMA · CIFAR-10 | ASR (%): 26.0% with 2 pixels to 69.5% with 50 pixels | Improved |
| increase number of perturbed pixels: effect on time (ablation) | FastJSMA · CIFAR-10 | time (s): 0.0003-0.0004 seconds | No clear effect |
Closest papers by meaning
- Backdoor Attacks on the DNN Interpretation System (AAAI 2022 · similarity 0.85)
- On the Connection Between Adversarial Robustness and Saliency Map Interpretability (ICML 2019 · similarity 0.85)
- Structured Gradient-based Interpretations via Norm-Regularized Adversarial Training (CVPR 2024 · similarity 0.83)
- Towards Understanding and Improving Adversarial Robustness of Vision Transformers (CVPR 2024 · similarity 0.83)
- Stop Walking in Circles! Bailing Out Early in Projected Gradient Descent (CVPR 2025 · similarity 0.83)
- Saliuitl: Ensemble Salience Guided Recovery of Adversarial Patches against CNNs (CVPR 2025 · 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.
