Konsep Inti
Proposing PiGW as a versatile framework for integrating watermarks into generative images, ensuring invisibility and robustness.
Abstrak
Abstract
PiGW introduces a framework for watermarking generative images.
Embeds watermarks into initial noise using adaptive frequency spectrum masks.
Demonstrates true invisibility and resistance to noise attacks.
Introduction
Latest generative models enable creation of indistinguishable images.
Urgent need for copyright protection tools with the rise of Artificial Intelligence Generated Content (AIGC).
Post-hoc Watermarking vs. Generative Watermarking
Traditional vs. generative watermarking methods explained.
Generative watermarking embeds watermarks during image generation for true invisibility.
Method
PiGW framework consists of 4 modules: Embedding, Generation, Attack, Authentication.
Embedded Module encodes key and combines with noise to create watermark embedding vector.
Experiments
Evaluation metrics include FID, CLIP Score, AUC, TPR@1%FPR.
Results show strong robustness of PiGW against various attacks and compression algorithms.
Conclusion
Plugin-based watermarking framework seamlessly integrates into existing generative models.
Achieves true invisibility and high robustness across different modalities.
Statistik
"The project code will be made available on GitHub."
"Image size set to 512x512."
"Watermark fixed at 30 bits."