Implementation of adversarial attack on different deep NN classifiers, the attacks are base on the algorithms in the papers :

ATTACKS :

• LBFGS Attack : Explaining and harnessing adversarial examples.
• FGSM : Intriguing properties of neural networks.
• Vanilla Attack :

DEFENSES :

• Distilled neural network : Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks
• Adversarial Training
• Binary thresholding

├── Adversarial_blackbox_attacks.ipynb  <- Notebook for testing BB attacks>
├── Adversarial_whitebox_attacks.ipynb  <- Notebook for testing WB attacks>
├── attack.py                           <- Test functions of attacks>
├── Attacks                             
│   ├── FGSM.py                         <- FGSM fast attack class>
│   ├── LBFGS.py                        <- L-BFGS attack class >
│   └── VanillaGradient.py              <- Vanilla attack class >      
├── Defense.ipynb                       <- Notebook for testing defenses>
├── defense.py                          <- Test functions of defense>
├── imagenet_classes.txt
├── Net.py                              <- Architectures of models >
├── Results                             <- Resulting images and accuracies >
├── utils.py                            <- Plotting functions >
└── weights                             <- Weights for pretrained models>

Our work was inspired by Adversarial Attacks and Defences Competition, we implemented 3 differents attack vectors and 3 matching defenses.

• Adversarial_whitebox_attacks.ipynb : We first implemented the attacks on the architecture Net.py with MNIST dataset, the notebook show the impact of our different attacks on the accuracy of the model
  • 
• Defense.ipynb : This notebook showcases the robustness of 3 different defenses against the attacks. You'll find the accuracy measure of the model when adding the defense. The L-BFGS attack was left out of the testing because the high computational cost of the attack.
  • 
• Adversarial_blackbox_attacks.ipynb : One very interesting feature of adversarial examples is their ability to transfer to different models. We tested this unique property by attack a model based on image generated froma different one. We used a more complex dataset (ants/bees) of 3 channels images from this test.
  • 
  • Note the dataset is available to download by running :

        wget https://download.pytorch.org/tutorial/hymenoptera_data.zip