evasion-and-preprocessing
Use ART evasion attacks, preprocessing defences, and adversarial training recipes for image/tabular robustness workflows.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- evasion-and-preprocessing
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Use ART evasion attacks, preprocessing defences, and adversarial training recipes for image/tabular robustness workflows.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/attack-api-reference.md10.7 KB
- 参考文档references/preprocessing-and-training.md9.6 KB
- 参考文档references/troubleshooting.md7.0 KB
03 SKILL.md 正文
Evasion and preprocessing
Use this sub-skill when the task is to generate evasion adversarial examples, select a white-box/black-box/physical image attack, add preprocessing defences, or plan adversarial training with ART estimators.
Route here
- Craft FGSM/FGM, PGD/BIM/MIM/AutoPGD/AutoAttack, Carlini, DeepFool, ElasticNet, JSMA, universal perturbation, virtual adversarial, HopSkipJump, Boundary, ZOO, SimBA, Square, Pixel/Threshold, SignOPT, DecisionTree, spatial, feature-adversary, or image-patch attacks.
- Add image/tabular preprocessing defences such as standardisation, spatial smoothing, feature squeezing, JPEG compression, Gaussian augmentation, label smoothing, Mixup, CutMix, Cutout, thermometer encoding, total variation minimization, or PixelDefend.
- Plan adversarial training with
AdversarialTrainer,AdversarialTrainerMadryPGD, or TRADES-style trainers.
Route away
- Model wrapping,
clip_values,input_shape, backend device setup, and estimator construction details belong to../estimators-and-models/or../setup-and-backends/. - Poisoning, backdoors, membership/attribute inference, model inversion, or model extraction belong to
../poisoning-inference-extraction/. - Robustness metrics, certification, tree verification, security curves, and gradient checks after attack generation belong to
../evaluation-and-certification/. - Speech, object-detection-heavy, malware, and audio-heavy evasion workflows are outside this selected runtime scope unless a future refresh adds their required backends. Recognize their ART class names, but do not promise runnable coverage here.
Operating sequence
- Confirm the estimator capability: loss gradients for PGD/FGM/BIM/MIM/AutoPGD, class gradients for Carlini/DeepFool/JSMA-style attacks, prediction-only classifiers for hard-label black-box attacks, or neural-network/image-specific support for patch/spatial attacks.
- Match all perturbation budgets to the estimator input scale and
clip_values; for normalized[0, 1]images use budgets such as8/255, not8. - Check shape conventions before running an attack: PyTorch image estimators usually use
channels_first=TrueandNCHW; many NumPy image preprocessors default tochannels_first=FalseandNHWC. - For targeted attacks, pass target labels as
ytogenerate; use one-hot labels unless the wrapped estimator workflow explicitly uses class-index labels. - Add preprocessing defences either directly as callable preprocessors or through estimator
preprocessing_defences; do not treat preprocessing alone as proof of robustness. - For adversarial training, start from a bounded PGD/FGM attack, verify labels and clipping, then choose generic, Madry PGD, or TRADES training according to the estimator backend.
References
- Attack family/API chooser: references/attack-api-reference.md
- Preprocessing and adversarial training recipes: references/preprocessing-and-training.md
- Evasion/preprocessing troubleshooting: references/troubleshooting.md
Bundled checks
- Tiny CPU PyTorch PGD/FGM smoke: scripts/smoke_evasion_pytorch.py
- Tiny NumPy standardisation/spatial smoothing smoke: scripts/smoke_preprocessor_numpy.py
Run bundled scripts with --help first; they use synthetic data and do not download datasets.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.