Adversarial Robustness

Description

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:

Repo structure

• data/ training data and adversarial perturbations
• notebooks/
• results/ collected results and plots
  • images/
• src/ implementations
  • RandomProjections/ methods based on random projections
  • BayesianSGD/ 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.