Command
vraven adversarial
Run finite-budget empirical adversarial stress testing
Minimal form
vraven adversarial <model>Arguments and options
| Option | Meaning | Default |
|---|---|---|
model | PyTorch model file | None |
--model-format | Model storage format. VRAVEN instrumentation supports state_dict and trusted pickle models; TorchScript can be loaded but cannot expose PyTorch hooks. | state_dict |
--model-factory, --factory | Import path such as my_project.models:Classifier | None |
--factory-kwargs | JSON object passed to the model factory | {} |
--state-dict-key | — | None |
--strip-prefix | — | None |
--device | — | cpu |
--non-strict | — | False |
--trusted-model | Allow Python-pickle model loading | False |
--input | Input tensor file | None |
--labels | Optional labels for the input batch | None |
--scope | Audit one sample, fixed-epsilon batch accuracy, or an epsilon robustness curve | sample |
--sample-index | — | 0 |
--target | Decision target; defaults to label or prediction | None |
--model-name | — | None |
--output-path | Dot path to logits in structured output | None |
--attacks | quick, full, or comma-separated attacks such as fgsm,pgd_linf,deepfool | quick |
--epsilon-grid | — | 0.01,0.05,0.1 |
--epsilon | — | 0.1 |
--protocol | — | vraven_ensemble_lite |
--max-samples | — | None |
--sampling | — | head |
--steps | — | 10 |
--restarts | — | 1 |
--step-size | — | None |
--clip-min | — | None |
--clip-max | — | None |
--seed | — | 24032026 |
--deepfool-iters | — | 40 |
--square-queries | — | 80 |
--fab-iters | — | 20 |
--fab-restarts | — | 2 |
--target-classes | — | 3 |
--format | — | json |
--output | Output path; defaults to the selected format | None |
Exact local helpRun
vraven adversarial --help against your installed version before automating a release workflow.