Research map / Adversarial robustness and security / Certified and empirical robustness
Sparse-PGD: A Unified Framework for Sparse Adversarial Perturbations Generation
Xuyang Zhong, Chen Liu
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 decomposition delta = p*m to PGD0 without restart (ablation) | PGD0 · CIFAR-10 Fast-EG-l1 | robust accuracy: 49.4 -> 58.0 | Hurt |
| add continuous mask and l0 projection (ablation) | sPGD_proj · CIFAR-10 Fast-EG-l1 | robust accuracy: 58.0 -> 33.9 | Improved |
| add random reinitialization (ablation) | sPGD_proj · CIFAR-10 Fast-EG-l1 | robust accuracy: 33.9 -> 18.1 | Improved |
| vary step size alpha for magnitude from 1/16 to 1 (ablation) | sPGD_proj · CIFAR-10 Fast-EG-l1 | robust accuracy: 19.6 -> 18.1 -> 19.0 | No clear effect |
| vary attack reinitialization tolerance t from 1 to 10 (ablation) | sPGD_proj · CIFAR-10 Fast-EG-l1 | robust accuracy: 18.1 -> 18.5 | No clear effect |
| random backward function policy instead of always projected (ablation) | sTRADES · CIFAR-10 sAA | robust accuracy: 41.5 -> 61.7 | Improved |
| alternate backward function every 5 epochs instead of random (ablation) | sTRADES · CIFAR-10 sAA | robust accuracy: 61.7 -> 51.5 | Hurt |
| train with 6x larger epsilon than test (multi-epsilon) (ablation) | sTRADES · CIFAR-10 sAA | robust accuracy: 34.0 -> 61.7 | Improved |
| increase multi-epsilon from 6x to 10x (ablation) | sTRADES · CIFAR-10 sAA | robust accuracy: 61.7 -> 55.5 | Hurt |
| raise training reinitialization tolerance from 3 to 10 (ablation) | sTRADES · CIFAR-10 sAA | robust accuracy: 51.7 -> 61.7 | Improved |
| multi-epsilon instead of multi-size strategy for 3x3 patch training (ablation) | sTRADES · CIFAR-10 3x3 patch | sAA robust accuracy: 67.9 -> 72.3 | Improved |
| use sTRADES instead of sAT | sPGD adversarial training · CIFAR-10 eps=20 | sAA robust accuracy: 36.2 -> 61.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SparseFool — SparseFool: A Few Pixels Make a Big Difference (CVPR 2019) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations (ICML 2024 · similarity 0.97)
- Sparse Adversarial Attack via Perturbation Factorization (ECCV 2020 · similarity 0.92)
- BRUSLEATTACK: A QUERY-EFFICIENT SCORE- BASED BLACK-BOX SPARSE ADVERSARIAL ATTACK (ICLR 2024 · similarity 0.88)
- Transferable Structural Sparse Adversarial Attack Via Exact Group Sparsity Training (CVPR 2024 · similarity 0.88)
- $\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples (ICLR 2025 · similarity 0.88)
- GSE: Group-wise Sparse and Explainable Adversarial Attacks (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.
