Stress-test a model under a declared budget.

Measure empirical behaviour under controlled attacks without presenting the result as a formal guarantee.

Run a quick sample audit

vraven adversarial model.pt \
  --model-format state_dict \
  --factory my_project.models:Classifier \
  --input input.npy \
  --scope sample \
  --attacks quick \
  --epsilon 0.03 \
  --format html \
  --output adversarial-report.html

Measure an epsilon curve

vraven adversarial model.pt \
  --model-format state_dict \
  --factory my_project.models:Classifier \
  --input batch.npy \
  --labels labels.npy \
  --scope curve \
  --attacks fgsm,pgd_linf \
  --epsilon-grid 0.0,0.01,0.02,0.03 \
  --max-samples 128 \
  --format html \
  --output robustness-curve.html

• Record the norm, epsilon, steps, restarts and clipping range.
• Use representative samples and supply labels for accuracy curves.
• Treat attack failure as “not found within this budget”, not proof of robustness.