Adversarial Attacks on Images by Relighting

Find artificial relightings that fool classifiers with PyTorch.

Project Description

• perform research on the field of available image augmentation techniques that simulate different illumination
• adapt the most suitable frameworks to attack the classifier
• evaluate the robustness of the targeted model against this type of input pertubations
• analyze whether the adversarial training approach helps improving its robustness.

Repo Structure

Attacks: this package contains all our adversarial attack algorithms
Classifiers: this package contains all classifiers that we are attacking in our experiments
Data: contains data we train the classifiers on
Dep: deprecated notebooks and scripts
Experiments: notebooks that execute our adversarial attack experiments
Relighters: model implementations of all relighters

Contributors

• Johanna
• Pavel
• Andreea
• Thank you to our advisor Aleksei!