NeuralNetworkNuke

Python tool to visualise attacking and defending neural networks

GUI

• GUI built using streamlit
• To use app locally,
  • Clone the repo and run StartUI.sh to view the app on your browser!
• App is also hosted remotely on heroku using my common host app,

  • https://infinityjoker-apps.herokuapp.com/

  • In the Common Host App, simply choose a project to load and click load and deploy.

  • Then go ahead and use the app! :)

  • If you want to change to another app, simply click on View Other Projects in top left and choose any other project and load and deploy.

Attack Neural Networks

• FGSM : Normal, Targeted, Iterative
• Jacobian Based Saliency Map Attack

Defend Neural Networks

Adversarial Training

• Train on the adversarial examples generated by different attacks on top of the model to make it robust to those attacks
• Unreliable as the attacks may still be able to fool the model
• Takes extra training time

Barrage of Random Transforms

• Train model on a barrage of transformed input images
• Much more robust than adversarial training
• Still unreliable and takes extra training time