contributor_README.md

HyperForensics++ Dataset Contribution Guide

This guide outlines the steps to contribute to the HyperForensics++ dataset repository.

---

1. Clone the Repository

First, clone the repository from Hugging Face to your local machine:

# Make sure git-lfs is installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/datasets/OtoroLin/HyperForensics-plus-plus
cd ./HyperForensics-plus-plus

2. Extract the Dataset

Decompress the .tar.gz files into a separate local hierarchy. You can utilize the provided unzipping.sh script. First, modify the ROOT_DIR variable to the local_data path

# In zipping.sh
# Defined the root directory of the whole dataset
ROOT_DIR="/root/to/local_data"

Due to the redundancy of our dataset, we use pigz program instead of standar gipz program to compress the files. Make sure you install it first.

sudo apt install pigz
# If you are using miniconda
conda install pigz

2.1 Decompressing the specify method and configuration:

./unzipping.sh --method <method name> --config <config name>

Replace <method name> with the forgery method and <config name> with the configuration index. This will automatically decompressed .tar.gz file and put it under hierarchy of the local dataset directory. For instance, if I want to decompress the images under ADMM-ADAM forgery method, config0 configuration:

./unzipping.sh --method ADMM-ADAM --config 0

This will decompress HyperForensics-plus-plus/data/ADMM-ADAM/config0/config0.tar.gz to /path/to/local/ADMM-ADAM/config0/*.

2.2 Decompress the entire hierarchy

./unzipping.sh --all

This will recursively decompress all tar.gz file in the HyperForensics-plus-plus/data and automatically put the decompressed files under hierarchy of the local dataset directory.

3. Create a pull request

The pull requests on Hugging Face do not use forks and branches, but instead custom “branches” called refs that are stored directly on the source repo.

The advantage of using custom refs (like refs/pr/42 for instance) instead of branches is that they’re not fetched (by default) by people cloning the repo, but they can still be fetched on demand.

3.1 Create a new PR on the hugging face

Go to community $\to$ New pull request $\to$ On your machine $\to$ enter Pull request title $\to$ Creat PR branch

Then, you shall see a brand new PR page with the title of your PR on the top left. Make note of the index number next to the title #{index}.

In this step, you actually create a new custom branch on remote under reference of ref/pr/#{index}.

The PR is still in draft mode, the maintainer can not merge the PR untill you publish the PR.

3.2 Create a new PR branch on local

git fetch origin refs/pr/<index>
git checkout pr/<index>

Fetch the remote PR branch ref/pr/<index> to local PR branch. Then checkout to the newly created custom branch under references pr/, and just for unambiguity, using the PR index as the name of that reference. You can actually create a new branch locally here as usual.

3.3 Push the PR branch

After you finish your modification, push the local PR branch to remote Hugging Face. Check it out here at 6.

4. Modify or Generate Data

Make changes or generate new data in the extracted directory. Ensure the data follows the required structure:

local_data/
    |-{attacking method}/
    |   |-config0/
    |   |   |-images.npy
    |   |-config1/
    |   | ...
    |   |-config4/
    |-{attacking method}/
    |   |-config{0-4}/
    |-Origin
    |   |-images.npy
HyperForensics-plus-plus/
    |-data/

5. Zip the Directory

After modifying or generating data, zip the directory into .tar.gz files and place it in the repository. You can utilize the provided zipping.sh script. First, modify the ROOT_DIR variable to the local_data path

# In zipping.sh
# Defined the root directory of the whole dataset
ROOT_DIR="/root/to/local_data"

Due to the redundancy of our dataset, we use pigz program instead of standar gipz program to compress the files. Make sure you install it first.

sudo apt install pigz
# If you are using miniconda
conda install pigz

5.1 Compressing the specify method and configuration:

./zipping.sh --method <method name> --config <config name>

Replace with the forgery method and with the configuration index. This will automatically put the compressed .tar.gz file under hierarchy of the dataset repo. For instance, if I want to zip the images under ADMM-ADAM forgery method, config0 configuration:

./zipping.sh --method ADMM-ADAM --config 0

This will compress /path/to/local/ADMM-ADAM/config0/* to HyperForensics-plus-plus/data/ADMM-ADAM/config0/config0.tar.gz.

5.2 Compressing the specify direcotry

./zipping.sh --dir-path <path/to/directory>

Replace <path/to/directory> with the intended directory path. This will compress the directory and put it in the working directory. For instance,

./zipping.sh --dir-path /path/to/local/ADMM-ADAM/config0

This will compress /path/to/local/ADMM-ADAM/config0/* to HyperForensics-plus-plus/config0.tar.gz.

5.3 Compress the entire hierarchy

./zipping.sh --all

This will compress the entire local data hierarchy seperately and automatically put the compressed files under hierarchy of the dataset repo.

Note that I did not imiplement the method/config-specified compressing for Origin since it seems unlikely people needs to compress/decompress it regularly. You can either do it manually or use the path-specified compressing method and put it in to hierarchy by hand.

6. Create a Pull Request

Push your changes to created PR:

git push origin pr/<index>:refs/pr/<index>

Replace <index> with the pull request index. And click the Publish button on the buttom of the page.

You can aso leave comment for your PR to showcase where you have modified.

6. Wait for merging the Pull Request

Once the pull request is published, the repository maintainer will review and merge it.

---

Thank you for contributing to the HyperForensics++ dataset!

Divide my study into three main scope

1. The data generation
2. The quality matricies to determind the quality of our dataset
3. The forgery detection baseline models

The data generation

Removal Method

This scope has been almost done by min-zuo with two algorithm 1. ADMM-ADAM; 2. FasyHyIn.

However, since both strategy looks into spectral domain to inpaint the area. I have to look carefully to make sure it actually remove the data instead of recovering it. How to do that? I need a classifier or segmentation model to detect the image making sure it has a wrong output

Object replacement

This scope is under min-zuo's hand. He use CRS_diff diffusion model + harmonization to generate the replacemnt example. :::warning The images example above water are hard the current bottleneck. Needs more attention :::

Adv attack

This is implemented by 恩召, I did not actually ask much about it. Choose one adv-attack strategy. Focusing on highly transibale attack

Qaulity matrics

Spectral Consistency Metrics

These metrics are used in HSI-tampering, which are used to ensure that the forged region's spectrum matches the surrounding context

1. SAM
2. Spectral RMSE
3. Pearson correlation (need more info)
4. EGRAS (need more info)

Forgery Detectability Metrics

I think this is the most important one since we are making forgery dataset. We want to show:

• rate of forgery images being mis-classified is high
  1. FNR
  2. Attack Success Rate
  3. Detection Score Drop (detector confidence before/after forgery) We can use either a classification model or segmentation model
• rate of forgeries not detected by a detection model is low

Visual Inspection

1. User studies - using false color?
2. Visual result figures (origin vs. forged vs. mask)

Visual Quality Matrics

This is relatively not important since this is a forgery dataset, how similar the forgery image is to the ground truth can not evaluate the goodness of the forgery

1. PSNR
2. SSIM
3. LPIPS(Learned Perceptual Image Patch Similarity)

Forgery Detection Baseline Model

From previous paragraph Forgery Detectability Metrics. We need:

1. Pixel wise existing forgery detection model
2. Our own modern SOTA detection model for each attacking method