shrebox / proactive-and-reactive-measures-for-adversarial-defense

Maximally separating features in intermediate feature layers using PCL loss + image transformations with adversarial example transferability.

Proactive-and-Reactive-Measures-for-Adversarial-Defense

In this project we aim to study the effects of reactive and proactive attacks as an adversarial defense particularly ona base reactive defense (PCL as termed in paper refer) that maximally separates features in intermediate layers in a deep learning model. Also, we study the effects of image transformations on feature space and adversarial example transferability.