lfz / guided-denoise

The winning submission for NIPS 2017: Defense Against Adversarial Attack of team TSAIL

Guided-Denoise

The winning submission for NIPS 2017: Defense Against Adversarial Attack of team TSAIL

Paper

Defense against Adversarial Attacks Using High-Level Representation Guided Denoiser

File Description

• prepare_data.ipynb: generate dataset

• Originset, Originset_test: the folder for original image

• toolkit: the program running the attack in batch

• Attackset: the attacks

• Advset: the adversarial images

• checkpoints: the models checkpoint used, download here

• Exps: the defense model

• GD_train, PD_train: train the defense model using guided denoise or pixel denoise

How to use

the attacks are stored in folder Attackset the script is in the toolkit folder. in the run_attacks.sh file: modify models to the attacks you want to generate, separate by comma, or use "all" to include all attacks in Attackset. use the command to run:

bash run_attacks.sh $gpuids

where gpuids is the id of the gpus you want to use, they are number separated by comma. It will generate the training set. Then change the line DATASET_DIR="${parentdir}/Originset" to DATASET_DIR="${parentdir}/Originset_test", and run the command bash run_attacks.sh $gpuids again.

Then specify a model you want to use, the models are stored in Exp folder, there is a sample folder, it refers to a model named "sample", let's use it. Then go to GD_train if you want to use guided denoiser, run

python main --exp sample

The program will load Exp/sample/model.py as a model to train. and also you can specify other parameters defined in the GD_train/main.py

guided-denoise's Issues

Link for Imagenet subset

Can you provide the extracted images from Imagenet? The original dataset is too big to download.

Models Checkpoints Link issue

Hello,

The Models checkpoint link is not working correctly.
Could you please provide them in the Repo ?

Thanks

How to get the rate of defense?

Why num classes is set to 1001?

https://github.com/lfz/Guided-Denoise/blob/master/Exps/sample/inception.py#L329
https://github.com/lfz/Guided-Denoise/blob/master/Exps/sample/inception.py#L413
But actually you have 1000 classes, why set to 1001 and then outputs.append(x[:,1:])

Training Dataset Size

Hi,
I wonder how many images do you select per class as training dataset? In the paper, you select 30 images per class but in prepare_data n_per_class is set to 4.

Thanks

FGD and LGD

Exps文件夹下的inceptionv3_state.ckpt是已经训练好的模型吗

the load of dataset

in prepare_data.ipynb, dose the imagenet_path means the whole imagenet2012 datasets? it seems so big for this dataset

Meet problem when load model dict

Hello, I have met some problem with loading the dictionary of the model.

The code of mine is like this. I choose InceptionV3 model. I want to load the "inception_v3.ckpt" dictionary provided by you.
However, it failed and return an error. The error is shown below.

    input_size = [299, 299]
    block = Conv
    fwd_out = [64, 128, 256, 256, 256]
    num_fwd = [2, 3, 3, 3, 3]
    back_out = [64, 128, 256, 256]
    num_back = [2, 3, 3, 3]
    n = 1
    hard_mining = 0
    loss_norm = False
    denoise = Denoise(input_size[0], input_size[1], block, 3, fwd_out, num_fwd, back_out, num_back)
    net = Inception3(denoise)

    model_path = "inception_v3.ckpt"
    pretrain_dict = torch.load(model_path)
    state_dict = net.state_dict()
    for key in pretrain_dict.keys():
        assert state_dict.has_key(key)
        value = pretrain_dict[key]
        if not isinstance(value, torch.FloatTensor):
            value = value.data
        state_dict[key] = value

    net.load_state_dict(state_dict)

Could you please tell me where is the bug? The pytorch version used by me is 1.2.0.
Thank you very much!

denoised model checkpoints

What the meaning of ‘requires_control’ in you code?

Hello!
I'm reading your code,but I can't understand the function of 'requires_control' flags in your code.
Could you please tell me?
Thank you!

from importlib import input_module

where is your model define file?

Which version of tensorflow did you use in this respository?

When i tried to run this code, i encountered the following errors in the 'attack_iter.py' file.

W tensorflow/core/framework/op_kernel.cc:1192] Invalid argument: Shape mismatch in tuple component 0. Expected [299,299,3], got [2038,1500,3]

OutOfRangeError (see above for traceback): FIFOQueue '_1_batch/fifo_queue' is closed and has insufficient elements (requested 3, current size 0)
         [[Node: batch = QueueDequeueUpToV2[component_types=[DT_UINT8, DT_STRING], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/device:CPU:0"](batch/fifo_queue, batch/n)]]

I just changed the image format from '.png' to '.jpg', and changed the following code:

    image = tf.image.decode_png(image_file)

to

    image = tf.image.decode_jpeg(image_file, channels=3)

I have tried some methods to solve it, but nothing worked. So i guess maybe the version of my tensorflow(1.4) is different from yours.

Wish to receive your reply, thank you!