This respository gives some test results of adversarial examples generation algorithms such as FGSM, CW, Deepfool and JSMA ,and we also print output of the last layer to exploit possible adversarial example detecting method.
If you want to test these projects, following dependencies you must need:
- python3.5
- tensorflow
Installing TensorFlow will take care of all other dependencies like numpy，keras and scipy.
The target models of this algorithms has been pretrained, by programing:
models.resnet34(pretrained=True)return a pretrained models with ImageNet datasets. Or you can train a model yourself. In our work, we train a classification models with MNIST datasets.
This part we cite the work of Papernot et al.. Default model in the source code is a deep neural network defined in above respository. This part relies on cleverhans's other files, you my need to install the whole respository for running this code.
We also cite this work from cleverhans.This tutorial covers how to train a MNIST/CIFAR model using TensorFlow, craft adversarial examples using the fast gradient sign method, and make the model more robust to adversarial examples using adversarial training.
CW attack consists of L0 attack,L2 attack and Li attack. In our work, we only test L2 attack.This tutorial covers how to train a MNIST model using TensorFlow, craft adversarial examples using CW attack. More details in C&W attack.
Min F, Qiu X, Wu F. Adversarial attack? Don't panic[C]//2018 4th International Conference on Big Data Computing and Communications (BIGCOM). IEEE, 2018: 90-95.
[1]. Goodfellow I J, Shlens J, Szegedy C. Explaining and Harnessing Adversarial Examples[J]. Computer Science, 2014.
[2]. Kurakin A, Goodfellow I, Bengio S. Adversarial examples in the physical world[J]. 2016.
[3].Szegedy, Christian, Zaremba, Wojciech, Sutskever, Ilya, Bruna, Joan, Erhan, Dumitru, Goodfellow, Ian J., and Fergus, Rob. Intriguing properties of neural networks.CoRR, abs/1312.6199, 2013.
[4]. guyen, A., J. Yosinski, and J. Clune, Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. 2015: p. 427-436.
[5]. Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Pascal Frossard. “DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks”. Conference on Computer Vision and Pattern Recognition (CVPR) 2016.
[6].Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. “The Limitations of Deep Learning in Adversarial Settings”. IEEE European Symposium on Security and Privacy (Euro S&P) 2016.
[7]. Mopuri K R, Garg U, Babu R V. Fast Feature Fool: A data independent approach to universal adversarial perturbations[J]. 2017.
[8]. Nicholas Carlini and David Wagner. 2017. Towards Evaluating the Robustness of Neural Networks. In IEEE Symposium on Security and Privacy (Oakland) 2017.