hidden-trigger-backdoor-attack

An implementation of the hidden trigger backdoor attack introduced by Saha et al.

Note: This repo's content is based on the main repo of the paper which can be found here: https://github.com/UMBCvision/Hidden-Trigger-Backdoor-Attacks

about the files and variables (the original version)

├── {data_root} variable, a.k.a. {clean_data_root} (the "data" folder)
│   ├── train
│   │   ├── n0123
│   │   │   ├── n0123_3451.JPEG
│   │   │   ├── n0123_6769.JPEG
│   │   │   ├── n0123_4341.JPEG
│   │   ├── n0234
│   │   ├── n0345
│   ├── val
│   ├── test
│   ├── triggers
│   │   ├── trigger_10.png
│   │   ├── trigger_11.png
│   │   ├── ...
│   │   ├── trigger_19.png
│   ├── {exp_ID} (will be created)
│   │   ├── multi_source_filelist.txt (WBC)
│   │   │   ├── LINE 1: n0123/n0123_3401.JPEG
│   │   │   ├── LINE 2: n0234/n0234_6941.JPEG
│   │   ├── finetune_filelist.txt (WBC)
│   │   ├── test_filelist.txt (WBC)
│   │   ├── patched_filelist.txt (WBC)
│   │   ├── poison_filelist.txt (WBC)

├── ImageNet_data_list
│   ├── finetune
│   │   ├── n0123.txt
│   │   │   ├── LINE 1: n0123/n0123_3401.JPEG
│   │   ├── n0234.txt
│   │   │   ├── LINE 1: n0234/n0234_6981.JPEG
│   ├── poison_generation
│   │   ├── n0345.txt
│   ├── test
│   │   ├── n0345.txt

├── patched_data/{exp_ID}/rand_loc_false/eps_16/patch_size_30/trigger_10/
│   │   ├── patched_data/arman/rand_loc_True/eps_16/patch_size_30/trigger_10/badnet_n03461385_37_epoch_0000054.png

├── poison_data/{exp_ID}/rand_loc_false/eps_16/patch_size_30/trigger_10/
│   │   ├── poison_data/arman/rand_loc_True/eps_16/patch_size_30/trigger_10/loss_03395_epoch_01_n02437312_2125_n03461385_36144_kk_00168.png

data_root: This is actually a string variable in the code showing the folder containing the real images of the ImageNet dataset.
- In my case, the data_root was /Users/armanmalekzadeh/Documents/GitHub/hidden-trigger-backdoor-attack/data
- It contains three folders: train, val, and test.
- Inside each folder, there are folders with names called wnid. The wnid is something like n0123 and represents a class of objects. For instance, a unique wnid is associated with chicken, while another one represents cat.
- A string like {data_root}/train/n0123/n0123_30.JPEG is the path to a JPEG file of the class n0123 which should be used for the training phase.
data/triggers: A folder containing the triggers
- Each trigger is a 80x80 RGB image formatted as png
data/{exp_ID}/multi_source_filelist.txt: A file containing relative paths to the source class images.
- Each line of the file contains the relative path to an image belonging to a different source class.
  - For instance, the first line could be n0123/n0123_3401.JPEG while the second could be n0234/n0234/6951.JPEG.
ImageNet_data_list: A folder located in the root of the project containing some files showing the path to all images used for training, testing, and finetuning.
- It contains three folders: poison_generation, test, and finetune.
- Each of these folders contain files like n0123.txt.
- Inside a file like n0123.txt, we find the path to a JPEG file at each line.
  - For instance: n0123.txt contains a line like n0123/n0123_4291.JPEG.
source_wnid_list.txt: A file containing the source class IDs (remember, we wanted the images belonging to the source class to be classified as the ones belonging to the target class).
- We can consider multiple source classes (more than one class).
- Each source class ID should be written in one line of this file (source_wnid_list.txt).
- Therefore, this file contains lines like this:
  - n0123
  - n0234
  - n0345

arm-on/hidden-trigger-backdoor-attack

hidden-trigger-backdoor-attack

about the files and variables (the original version)