This repository consists of my implementation of Adversarial CNN model in pytorch, hitting 89.22% on Fashion-MNIST test dataset. Trained model also was provided in order to run the attack. Based on our trained model we implemented evasion attack (FGSM). Results are:
Epsilon: 0 Test Accuracy = 8922 / 10000 = 0.8922 /n
Epsilon: 0.05 Test Accuracy = 3595 / 10000 = 0.3595
Epsilon: 0.1 Test Accuracy = 1663 / 10000 = 0.1663
Epsilon: 0.15 Test Accuracy = 784 / 10000 = 0.0784
Epsilon: 0.2 Test Accuracy = 361 / 10000 = 0.0361
Epsilon: 0.25 Test Accuracy = 230 / 10000 = 0.023
Epsilon: 0.3 Test Accuracy = 181 / 10000 = 0.0181
In folder 'saved' you can see the result figure 'accuracy vs epsilon':
where the higher epsilon is, the lower the accuracy of our model. Next I would like to read papers where CNN model accuracy is higher than 90%, retrain my model and rerun again the attack.