This repository is the Python implementation of the "DefenseLayer", an intra-model defense layer approach to securing deep learning image classifiers against adversarial attacks. The paper describing our approach, titled "Defending Against Adversarial Attacks One Layer at a Time", is included in this repository.
Project presentation can be found: here
- Extracts an ImageNet test dataset composed of two classes: bikes and ships.
- Creates FGSM and DeepFool attacks based on the ImageNet dataset.
- Inserts the wavelet denoising layers into the model
- Tests the modifed models on the various test datasets
- Generates the graphs included in the paper
- jkurian49 - (https://github.com/jkurian49)