WIBE

WIBE is an extensible open source framework for evaluating imperceptibility and robustness of digital watermarks for generated images. The framework supports watermarking methods both during the image generation process and post-generation. So it is possible to evaluate watermarking techniques and attacks against them on both generated and original images.

The framework architecture consists of a core module and a set of extensions. The core implements Pipeline orchestration functionality. The extensions implement:

• Image watermarking Algorithms: currently 34 supported, including Watermark Anything, TrustMark, StegaStamp, Tree-Rings Watermarks, VideoSeal

• Attacks on watermarks: both common distortions and 30+ advanced attacks supported, including SADRE, DIP-based

• Datasets: DiffusionDB and MS COCO supported

• Metrics: 18 supported, including BER (Bit Error Rate), SSIM (Structural Similarity Index Measure), LPIPS (Learned Perceptual Image Patch Similarity), ImageReward

Contents

• Quick start
  • A. Clone the repository and navigate to its directory:
  • B. Configure the environment
  • C. HuggingFace
  • D. All set!
  • E. Results
  • HuggingFace Authentication Setup
• Manual setup
• System requirements
• wibench-venv: virtual environment manager
  • File layout
  • Commands
  • Per‑profile Python version
  • How the pipeline picks a venv
  • Practical notes
• Pipeline
  • Configuration file
• Algorithms
  • How to implement a new watermarking algorithm
  • Implemented algorithms
• Attacks
  • How to implement a new attack
  • Implemented attacks
• Datasets
  • How to add a new dataset
  • Implemented datasets
• Metrics
  • How to implement a new metrics
  • Implemented metrics