ZK-WAGON: Imperceptible Watermark for Image Generation Models using ZK-SNARKs

Watermarks can prove who created an AI image without revealing trade secrets: ZK-WAGON embeds invisible proof of origin into generated images, allowing creators to verify authenticity while keeping their model weights, prompts, and other sensitive information completely private—addressing a major barrier to adoption of previous watermarking methods.

The watermark is nearly impossible to detect or remove: By embedding the proof using imperceptible techniques (LSB steganography), the watermark doesn't degrade image quality like traditional methods and cannot be easily stripped away, making it practical for real-world deployment against misuse.

This approach works across different types of AI image generators: The system functions with both GAN and Diffusion models, suggesting it could become a broadly applicable standard for authenticating synthetic images rather than requiring custom solutions for each model type.

Speed improvements make the technology scalable: The paper's Selective Layer ZK-Circuit Creation method significantly reduces the time needed to generate proofs, making watermarking feasible for large-scale commercial use rather than remaining a theoretical exercise.

This addresses critical harms from synthetic media: By enabling verifiable proof of origin, ZK-WAGON provides a technical tool to combat misinformation, deepfakes, and intellectual property theft—three major societal risks highlighted as increasingly urgent as image generation becomes more accessible.