eth-sri/diffai

A certifiable defense against adversarial examples by training neural networks to be provably robust