Interactive demo
Watch a model break.
Pick an attack and a budget and see what happens to a 99%-accurate digit classifier and a YOLOv5 object detector. These are real results from ptAI's engine on its built-in models — pre-computed, so it's instant and nothing runs on your data.
1 · Attack
2 · Budget
Clean accuracy—
Under attack—
Digits fooled (below)—of 10
Avg. change per pixel (L∞)—
Accuracy vs. budget (500 test images)
1 · Attack
2 · Budget
3 · Image
mAP@50 (4 images)—
Confident detections—
Objects that vanished—IoU ≥ 0.5, same class
Change per pixel (L∞)—
Clean
Adversarial
Your model next
Run the same attacks on the model you ship.
Upload a Keras, PyTorch, scikit-learn, gradient-boosting or ONNX model in the dashboard — or call the same engine from your CI through the API.