InvisMark: Invisible and Robust Watermarking for AI-Generated Image Provenance

Rui Xu, Mengya Hu, Deren Lei, Yaxi Li, David Lowe, Alex Gorevski, Mingyu Wang, Emily Ching, Alex DengWACV 2025

TL;DR

This paper introduces InvisMark, a novel watermarking technique designed to embed invisible and robust identifiers into AI-generated images. It addresses the growing need for content provenance by offering high imperceptibility, a significantly larger payload capacity (256 bits), and strong resilience against various image manipulations and attacks, making it a promising solution for verifying the authenticity of digital media.

What Problem Does This Paper Address?

The rapid advancement of generative AI (GenAI) has made it incredibly easy to create hyper-realistic images, including deepfakes. This poses a significant challenge to information integrity, as these images can be used for malicious purposes like spreading misinformation or damaging reputations. While initiatives like C2PA aim to provide provenance information, metadata can be easily stripped. Traditional watermarking methods often struggle with high-resolution images, have limited payload capacity, and can introduce noticeable artifacts, hindering their adoption. This paper tackles the problem of creating a watermarking solution that is imperceptible, robust, and capable of carrying sufficient information to reliably track the origin of high-resolution AI-generated images.

Methods

InvisMark employs a novel approach that leverages advanced neural network architectures and training strategies. Key aspects include:

  • Resolution Scaling: The method incorporates resolution scaling during training, which helps the watermark decoder become more resilient to common image transformations without significantly impacting image quality.
  • MUNIT-based Encoder: A MUNIT-based encoder architecture with skip connections is used to embed the watermark, preserving fine image details.
  • ConvNeXT Decoder: A ConvNeXT-base network, known for its state-of-the-art performance in computer vision, is used as the decoder for recovering the watermark.
  • Robust Optimization: Instead of relying on randomly sampled noises for robustness, InvisMark treats watermark resilience as a robust optimization problem. It strategically focuses on worst-case scenarios during the final training stages to enhance performance against challenging distortions.
  • Larger Payload: The technique is designed to embed a 256-bit watermark, a significant increase over typical capacities, enabling the inclusion of identifiers like UUIDs with error correction codes for more reliable decoding.

Datasets

  • Training: The model was trained on 100,000 images generated by DALL-E 3 from randomly sampled user prompts.
  • Evaluation:
    • DALL-E 3 Dataset: Used for evaluation, consisting of 900 high-quality images.
  • DIV2K: A well-established super-resolution benchmark with 900 high-quality, 2K resolution images, used to evaluate performance on non-AI-generated content.

Key Insights and Ideas

  • High-resolution images have untapped capacity: The core insight is that high-resolution images inherently possess the capacity to embed a multitude of imperceptible signals, which InvisMark aims to harness.
  • Robust optimization is crucial for real-world resilience: Focusing training on worst-case scenarios, especially for geometric transformations, significantly improves the watermark's ability to withstand common image manipulations.
  • Larger payloads enable richer provenance: The ability to embed 256 bits is a game-changer, allowing for more complex identifiers and error correction, which directly translates to more reliable provenance tracking and fewer ID collisions.
  • Post-generation watermarking offers flexibility: Unlike methods that embed watermarks during the generation process, InvisMark can be applied to any image, regardless of its origin, making it broadly applicable.

Key Results

  • Imperceptibility: InvisMark achieves state-of-the-art imperceptibility, with PSNR values around 51 and SSIM around 0.998, significantly outperforming prior works that achieved around 40 PSNR.
  • Robustness: The method maintains over 97% bit accuracy across various image manipulations, with near-perfect decoding for many common noises.
  • Payload Capacity: InvisMark successfully embeds 256-bit watermarks while preserving image quality and robustness.
  • UUID Encoding: With 128-bit UUIDs and BCH error correction, InvisMark achieves a perfect 100% success rate for watermark decoding across various transformations, and over 90% even under challenging distortions.
  • Attack Resilience: The method shows good robustness against adversarial and regeneration attacks, though it becomes vulnerable at very low PSNR levels (below 25). It also demonstrates resilience against forgery attacks.

Limitations and Future Work

  • Subtle Artifacts: While generally imperceptible, InvisMark can, in rare instances, produce subtle artifacts in areas with large, uniform backgrounds (e.g., clear blue skies). Future work could focus on enhancing residual color uniformity to address this.
  • Vulnerability at Low PSNR: The watermark becomes highly vulnerable to removal when image quality degrades significantly (PSNR < 25).
  • Forgery Attack Mitigation: To address forgery attacks, the authors propose integrating InvisMark with complementary fingerprinting techniques (like Content Credentials) to bind watermarks to image content, enhancing security by rejecting mismatches.

Summary updated

BibTeX
@inproceedings{xu2025invismark,
  author = {Rui Xu and Mengya Hu and Deren Lei and Yaxi Li and David Lowe and Alex Gorevski and Mingyu Wang and Emily Ching and Alex Deng},
  title = {{InvisMark: Invisible and Robust Watermarking for AI-Generated Image Provenance}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure