Data-efficient-UAP

The official implementation code of Improving Transferability of Universal Adversarial Perturbation with Feature Disruption, which had been accepted in IEEE Transaction on Image Processing. The table of contents is as follows.

1. Settings: preparing training and test dataset

First, you should prepare the dataset (i.e., ImageNet) as following folder structure:

1.1 ImageNet & Downstream dataset

These two types of datasets should be arranged by each image placed in their category folder. News: we release the constructed ImageNet10k, we train our UAP on 2000 images randomly sampled from this.

-- ImageNet_folder
    -- train_folder
    	-- n01440764
        -- n01443537
        -- ...
        -- n15075141
    -- val_folder
    	-- n01440764
        -- n01443537
        -- ...
        -- n15075141

1.2 COCO & VOC & Mixed

These three types of datasets should be arranged by the image folder directly, as our method does not require the label. These three types of datasets should be arranged by the image folder directly, as our method does not require the label. Note that the mixed is composed of four datasets, so they should be assigned with a list of datapaths according to this dataset format.

-- COCO2014
  -- train2014
    -- image_01.jpg
    -- xxx

2. Training: train the UAP for different models

We provide the code and the corresponding description for the train the UAP

python main.py  \
    --exp_name "exp" \
    --model_name "resnet50" \
    --train_data_name "imagenet" \
    --test_data_name "imagenet" \
    --train_data_path "train data path" \
    --test_data_path "test data path" \
    --save_dir "checkpoints" \
    --eps 10. \
    --miu 0.1 \
    --eps_step 0.001 \
    --input_size 224 \
    --batch_size 50 \
    --epochs 50 \
    --num_works 8 \
    --nb_images 2000 \
    --feat_type "half_feat_and_grad" \
    --loss_type "abs" \
    --sort_type "channel_mean" \

Description

python main.py  \
    --exp_name "exp" \  # folder name to save the intermediate result
    --model_name "resnet50" \ # the victim model
    --train_data_name "imagenet" \  # the dataset used to train the UAP 
    --test_data_name "imagenet" \  # the dataset used to test the UAP
    --train_data_path "train data path" \  # the data path used to train the UAP
    --test_data_path "test data path" \  # the data path used to test the UAP
    --save_dir "checkpoints" \    # folder used to save the final trained results
    --eps 10. \  # the maximum modification of the UAP
    --miu 0.1 \  # the weight of momentum
    --eps_step 0.001 \  # the updation step
    --input_size 224 \  # image size
    --batch_size 50 \   # batch size
    --epochs 50 \      # max training epoch
    --num_works 8 \   
    --nb_images 2000 \  # number of training image
    --feat_type "half_feat_and_grad" \  # feature usage type
    --loss_type "abs" \    # loss function design type
    --sort_type "channel_mean" \  # feature importance calculation standard

3. Testing: test the trained UAP

During test stage, you need to assign the ImageNet validation path and the UAP files, and can get the evaluation result.

We provided our trained UAP files in:

Different in Training settings	Link
Model	here
Number of training samples	here
Dataset	here

Download or assign the UAP file in the uap_validation_imagenet.py

Run the following code to get the final result.

python uap_validation_imagenet.py

Note that, the evaluation of UAP on downstream task can be achieved by assigning the corresponding dataset folder. The code is as follows

python uap_validation_imagenet.py --is_downstream True

4. Downstream Tasks

We provide three manner to train the downstream task

From Scratch: We do not evalute the UAP's performance on these models
Fixed: Fixed the feature extractor and finetuned the classification head
Fullnet: Finetuned the feature extraction and classification head

Training detail refers to transfer_learning.py, using the follow code to perform training

python transfer_learning.py

winterwindwang/Data-efficient-UAP