AI_Security_training

Setup

run python ./C_W.py
env : anoconda, python-3.5, numpy, tensorflow

Attack : Carlini and Wagner L2

Proposed by Carlini and Wagner. It is an iterative, white box attack.

paper reference : https://arxiv.org/abs/1608.04644

Model

model data structure : Reference from Cleverhan. Definded in models.py, witch providing numerous hirachical tensor model.

3 layer CNN

'Conv2D0': <tf.Tensor 'add_7:0' shape=(?, 14, 14, 64) dtype=float32>,
'ReLU1': <tf.Tensor 'Relu:0' shape=(?, 14, 14, 64) dtype=float32>,
'Conv2D2': <tf.Tensor 'add_8:0' shape=(?, 5, 5, 128) dtype=float32>,
'ReLU3': <tf.Tensor 'Relu_1:0' shape=(?, 5, 5, 128) dtype=float32>,
'Conv2D4': <tf.Tensor 'add_9:0' shape=(?, 1, 1, 128) dtype=float32>,
'ReLU5': <tf.Tensor 'Relu_2:0' shape=(?, 1, 1, 128) dtype=float32>,
'Flatten6': <tf.Tensor 'Reshape:0' shape=(?, 128) dtype=float32>,
'logits': <tf.Tensor 'add_10:0' shape=(?, 10) dtype=float32>,
'probs': <tf.Tensor 'Softmax:0' shape=(?, 10) dtype=float32>

Problems & some conclution

< https://docs.google.com/document/d/1NhyEjP_6181k1pIojbDmpK0pSM5XeCmY3nqBjCFQChc/edit?usp=sharing >