Adversarial examples are special inputs to deep learning models, maliciously crafted to fool them into incorrect outputs. Even the state-of-the-art models are vulnerable to adversarial attacks, thus a lot of issues arise in many security fields of artificial intelligence. In this repo we aim at investigating techniques for training adversarially robust models.
Examples of adversarial perturbations:
data/training data and adversarial perturbationsnotebooks/results/collected results and plotsimages/
src/implementationsRandomProjections/methods based on random projectionsBayesianSGD/implementation of Bayesian SGD from Blei et al. (2017)BayesianInference/BNN training using VI and HMC
trained_models/baseline/randens/randreg/bnn/
tensorboard/
Scripts should be executed from src/ directory.