CW_Attack_on_MNIST

Implemention of cw attack on pytorch with corresponding MNIST model

MNIST model

Based on Towards Evaluating the Robustness of Neural Networks TABLE1 Consist of four convolution layer, two pooling layers, tow FC layers and ReLU. Notice: softmax shouldn't be put into model. What we need is the result from FC.That's why I use crossEntropyLoss as loss function.

Carlini and Wagner L2

L0,L2,L-inf be used for generating adversarial examples. In this repo, I choose L2 norm as the objective function.

def loss1_func(w,x,d,c):
    return torch.dist(x,(torch.tanh(w)*d+c),p=2)

You can define different objective function f in def f(output,tlab,target,k=0)

Blog(Written in Chinese)

CW Attack