Generating Adversarial Examples using Adv-GAN

This project is adapted from https://github.com/mathcbc/advGAN_pytorch, and it is a Pytorch implementation of the paper "Generating Adversarial Examples with Adversarial Networks" (Adv-GAN) on MNIST dataset.

See the complete report in:
(English) https://github.com/yegmor/Final_Project/blob/main/english_report/english_report.pdf
(Persian) https://github.com/yegmor/Final_Project/blob/main/Final%20Report.pdf, and presentation (Persian) in https://github.com/yegmor/Final_Project/blob/main/Presentation.pdf.

NOTE: This implementation is a little different from the paper, because a clipping trick has been added.

Run code

Install the required packages

pip install -r requirements.txt

Training the target model

python3 train_target_model.py

Training the AdvGAN

python3 train_advGAN_model.py

Testing adversarial examples using generator network of the trained AdvGAN

python3 test_adversarial_examples.py

Results

• The plots are available in results directory.
• The trained models are available in models directory.

Outputs

Target Network Architecture

MNIST_target_net(
(conv1): Conv2d(1, 32, kernel_size=(3, 3), stride=(1, 1))
(conv2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1))
(conv3): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1))
(conv4): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1))
(fc1): Linear(in_features=1024, out_features=200, bias=True)
(fc2): Linear(in_features=200, out_features=200, bias=True)
(logits): Linear(in_features=200, out_features=10, bias=True)
)

Evaluate the target model on the test set

test_num_correct: 9928 total test data: 10000
model loss on testing set testing set: 1.612297
model accuracy on testing settesting set: 0.992800

Evaluate the target model on generated adversarial examples generated from train set

Training set per-class accuracy:
[(0, 0.2026000337666723), (1, 0.059329575793533075), (2, 0.35246727089627394), (3, 0.326211058554885), (4, 0.15405682985279015), (5, 0.27670171555063644), (6, 0.08448800270361609), (7, 0.22346368715083798), (8, 0.6494616304905144), (9, 0.15128593040847202)]
Training set F1 score (micro): 0.002450
Training set F1 score (weighted): 0.002756
Training set Accuracy score: 0.245000
Training set attack success rate: 99.755000

Plots

• Generated adversarial examples

---

• Confusion Matrix for the corresponding outputs of the target network for the adversarial examplesgenerated from train set