furonghuang-lab/ETI_Competition_Data

ETI Competition Data: Collected from NeurIPS 2024 Competition "Erasing the Invisible"

Dataset Description

Dataset ID: furonghuang-lab/ETI_Competition_Data

This dataset comprises the complete data from the "Erasing the Invisible": The NeurIPS 2024 Competition on Stress Testing Image Watermarks. The competition aimed to systematically evaluate the resilience of state-of-the-art invisible image watermarking techniques against a wide array of removal attacks developed by participants worldwide.

The dataset includes:

1. Original Watermarked Images: The initial set of images embedded with various watermarking algorithms, used in both the Beige-box and Black-box competition tracks.
2. Participant Submission Evaluations: Detailed evaluation results for all ~2,700 submissions received. This includes per-submission, per-watermark attack success rates, comprehensive image quality scores, and overall competition scores.
3. (Note on Attacked Images): The actual attacked image files submitted by participants are also part of this release and can be accessed. This dataset card primarily describes the metadata and evaluation scores; refer to the "How to Use" section or the dataset loading script for details on accessing the corresponding attacked images for each submission.

This resource is invaluable for researchers and practitioners in digital watermarking, adversarial machine learning, image forensics, and content authenticity. It provides a large-scale benchmark for:

• Analyzing watermark vulnerabilities and successful attack strategies.
• Developing and testing more robust watermarking defenses.
• Benchmarking new watermark removal algorithms.
• Research into image quality assessment under adversarial manipulations.

Dataset Structure

The dataset is organized into four main subsets:

1. Beige_Track_Images

• Purpose: Contains the original 300 images used in the Beige-box track of the competition. In this track, participants were informed of the specific watermarking algorithm applied.
• Content: 300 images.
• Features:
  • image_index (int64): A unique identifier for the image.
  • watermarked_image (Image): The watermarked image file (relative path to image file in this dataset repo if preview not shown).
• Watermarks Used:
  • Gaussian Shading: Applied to 150 images generated by Stable Diffusion 2.1.
  • StegaStamp: Applied to 150 images generated by Flux.1 [dev].

2. Black_Track_Images

• Purpose: Contains the original 300 images used in the Black-box track, where the watermarking methods were kept confidential from participants.
• Content: 300 images.
• Features:
  • image_index (int64): A unique identifier for the image.
  • watermarked_image (Image): The watermarked image file (relative path to image file in this dataset repo if preview not shown).
• Watermarks Used: A mix of single and double watermarks:
  • Single Watermarks (50 images each): JigMark, PRC, StableSignature, Trufo.
  • Double Watermarks (50 images each): Gaussian Shading + JigMark, StableSignature + StegaStamp.

3. Beige_Track_Submissions

• Purpose: Contains detailed evaluation metadata and scores for the 1,072 valid submissions made to the Beige-box track.
• Content: 1,234 records (may include preliminary or invalidated submissions before final count if this number is from the raw schema).
• Features:
  • submission_id (string): Unique identifier for the submission.
  • submission_time (timestamp[us]): Timestamp of the submission.
  • gaussianshading (dict), stegastamp (dict): Dictionaries containing detailed evaluation results (e.g., detection scores/flags) for the respective watermarks on the images attacked in this submission. The exact structure of the dictionary should be inspected by the user.
  • IQM Scores (all dict): aesthetics, artifacts, clip_fid, legacy_fid (corresponds to FID), lpips, nmi, psnr, ssim. These dictionaries contain per-image quality metric scores comparing the attacked images to their originals.
  • performance (float64): Overall watermark detection performance score ($A$) for the submission.
  • quality (float64): Overall image quality degradation score ($Q$) for the submission.
  • score (float64): Final competition score ($\sqrt{Q^2 + A^2}$) for the submission.

4. Black_Track_Submissions

• Purpose: Contains detailed evaluation metadata and scores for the 1,650 valid submissions made to the Black-box track.
• Content: 1,116 records (may include preliminary or invalidated submissions before final count if this number is from the raw schema).
• Features:
  • submission_id (string): Unique identifier for the submission.
  • submission_time (timestamp[us]): Timestamp of the submission.
  • Watermark Scores (all dict): gaussianshading, jigmark, prc, stablesig, stegastamp, trufo. Dictionaries with evaluation results for these watermarks on attacked images.
  • IQM Scores (all dict): aesthetics, artifacts, clip_fid, legacy_fid (corresponds to FID), lpips, nmi, psnr, ssim.
  • performance (float64): Overall watermark detection performance score ($A$).
  • quality (float64): Overall image quality degradation score ($Q$).
  • score (float64): Final competition score ($\sqrt{Q^2 + A^2}$).

Evaluation Metrics Recap

Submissions were evaluated on two primary fronts:

1. Image Quality Degradation ($Q$):

   • Quantifies the visual distortion introduced by an attack compared to the original watermarked image.
   • Aggregated from eight distinct Image Quality Metrics (IQMs): PSNR, SSIM, NMI, FID (referred to as legacy_fid in schema), CLIP-FID, LPIPS, $\Delta$Aesthetics, and $\Delta$Artifacts.
   • Lower $Q$ is better (less degradation).

2. Watermark Detection Performance ($A$):

   • Measures how well the watermark is still detected after an attack.
   • Defined as $A = \text{TPR}@0.1%\text{FPR}$ (True Positive Rate at a 0.1% False Positive Rate).
   • Lower $A$ is better for attackers, indicating more successful watermark removal (i.e., the watermark is less detected post-attack).

3. Overall Score:

   • Calculated as the Euclidean distance to the origin in the $Q-A$ space: $\text{Score} = \sqrt{Q^2 + A^2}$.
   • Lower Score is better, representing an optimal balance between effective watermark removal and preservation of image quality.

For full details on metric calculation and normalization, please consult the competition's technical report.

How to Use This Dataset

You can load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load all subsets into a DatasetDict
dataset_dict = load_dataset("furonghuang-lab/ETI_Competition_Data")

# Or load a specific subset (e.g., Beige Track original images)
beige_track_images = load_dataset("furonghuang-lab/ETI_Competition_Data", name="Beige_Track_Images")
print("First Beige Track image entry:", beige_track_images['train'][0])

# Example: Load Black Track submission metadata
black_track_submissions = load_dataset("furonghuang-lab/ETI_Competition_Data", name="Black_Track_Submissions")
print("First Black Track submission entry:", black_track_submissions['train'][0])

(Note on accessing attacked image files for submissions: Please refer to the dataset loading script or documentation provided with this Hugging Face dataset for specific instructions on how to retrieve the actual attacked image files corresponding to each submission_id.)

Data Fields and Schema

The dataset schema is as follows:

================================================================================
                              DATASET SCHEMA SUMMARY
================================================================================

----------------------------------------
SUBSET: Beige_Track_Images
  Number of rows: 300
  Features:
    - image_index: Value(dtype='int64')
    - watermarked_image: Image()

----------------------------------------
SUBSET: Black_Track_Images
  Number of rows: 300
  Features:
    - image_index: Value(dtype='int64')
    - watermarked_image: Image()

----------------------------------------
SUBSET: Beige_Track_Submissions
  Number of rows: 1234 # Raw count from schema
  Features:
    - submission_id: Value(dtype='string')
    - submission_time: Value(dtype='timestamp[us]')
    - gaussianshading: dict
    - stegastamp: dict
    - aesthetics: dict
    - artifacts: dict
    - clip_fid: dict
    - legacy_fid: dict # Corresponds to FID used in the paper
    - lpips: dict
    - nmi: dict
    - psnr: dict
    - ssim: dict
    - performance: Value(dtype='float64') # Metric A
    - quality: Value(dtype='float64')    # Metric Q
    - score: Value(dtype='float64')      # Overall Score

----------------------------------------
SUBSET: Black_Track_Submissions
  Number of rows: 1116 # Raw count from schema
  Features:
    - submission_id: Value(dtype='string')
    - submission_time: Value(dtype='timestamp[us]')
    - gaussianshading: dict
    - jigmark: dict
    - prc: dict
    - stablesig: dict
    - stegastamp: dict
    - trufo: dict
    - aesthetics: dict
    - artifacts: dict
    - clip_fid: dict
    - legacy_fid: dict # Corresponds to FID used in the paper
    - lpips: dict
    - nmi: dict
    - psnr: dict
    - ssim: dict
    - performance: Value(dtype='float64') # Metric A
    - quality: Value(dtype='float64')    # Metric Q
    - score: Value(dtype='float64')      # Overall Score

================================================================================

Considerations for Using the Data

• Limitations: The dataset reflects the specific watermarking methods, generative models (Flux.1 [dev], Stable Diffusion 2.1), and attack conditions of the NeurIPS 2024 "Erasing the Invisible" competition. Findings may not generalize to all watermarks or attack scenarios. The competition primarily focused on watermark removal attacks.
• Licensing: The ETI Competition Data is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
• Responsible Use: Users are encouraged to use this dataset responsibly, adhering to ethical AI practices, and for the advancement of research in content authenticity and security.

Citation

Coming soon.

Additional Information

• Competition Evaluation Program: https://github.com/erasinginvisible/eval-program
• Point of Contact: For questions about the dataset, please refer to the contact information in the associated technical report or open an issue on the GitHub repository for the evaluation program.