Fast Gradient Sign Method (FGSM) on a MNIST Handwritten Digit Classification Model

Adversarial examples generated via the FSGM method are used on a handwritten digit classification model with accuracy of more than 98%.

This project consists of two main python modules: train_model.py and attack_model.py.

• In train_model.py a classification model will be trained on the MNIST handwritten digit database. The model will also be saved for future use.
• In attack_model.py the trained model will be loaded and tested with adversarial examples generated with the FGSM.

Prerequisites

• Python: 3.7.4
  • Keras: 2.3.0
  • Tensorflow: 1.14.0
  • Cleverhans: 3.0.1

Results

The normal validation accuracy is: 0.9992857142857143
The adversarial validation accuracy is: 0.20952380952380953
The prediction for the original number is: [4]
The prediction for the adversarial example is: [8]
The prediction for the original number is: [1]
The prediction for the adversarial example is: [8]
The prediction for the original number is: [9]
The prediction for the adversarial example is: [8]