This repository contains source code for the main experiments of the NeurIPS 2022 paper titled 'What Can the Neural Tangent Kernel Tell Us About Adversarial Robustness?'.

robust_train_CNN.py contains code to adversarially train a neural network and compute (& save) empirical ntks during training. measurements.py contains code for the computation of several kernel quantities on precomputed ntks.

NTK_features_example.ipynb is a self-contained notebook (that can be run on google colab) that demonstrates the NTK features of an architecture, as defined in the paper.