FGSM
Test an image model with a single gradient-based adversarial perturbation.
- Testing group
- Robustness
- Applies to
- Adversarial attacks on image models
- App version
2.0.0- Channel
- stable
Required inputs
A model, evaluation data, a baseline result and the selected transformation or attack settings. Adversarial attacks require a compatible image model.
Test procedure
Establish the baseline, apply the selected perturbation and rerun the task metric under the changed conditions.
Interpreting results
Compare the perturbed result with the baseline and inspect affected examples. Judge the change against conditions plausible for your use case.
Limitations
Evidence record
Keep the baseline, perturbation configuration, changed examples and resulting performance comparison together.