Adversarial attack app! Want to confuse a neural network?

Want to confuse a neural network?

Try FGSM (Fast Gradient Sign Method) – a technique where adding carefully selected “noise” forces the model to misclassify what it “sees” in an image.

I wanted to observe what happens inside the neural network during such an attack, so I created an interactive visualization:

  1. Draw an apple or a pear – the model will try to recognize what’s in the image.
  2. Add “attack noise.”
  3. Watch what happens inside each layer.

Bonus: Try drawing an image that can resist this type of attack :slight_smile:

Link: https://attack-on-applepear.streamlit.app/

3 Likes