DiffMark Hides 256-Bit Keys in AI Images 17x Faster Than Rivals

DiffMark writes a 256-bit key into the last few denoising steps of a frozen diffusion model and reads it back in 89.4 ms, with no inversion.

·
·
·
DiffMark Hides 256-Bit Keys in AI Images 17x Faster Than RivalsPRO
  • DiffMark embeds arbitrary 256-bit keys into the late denoising steps of frozen diffusion models.
  • Reports 99.4–100% average TPR@1%FPR under attacks across SD 2.1, SD3 and FLUX on three prompt sets.
  • Extraction takes 89.4 ms with no diffusion inversion, about 17x faster than Tree-Ring and Gaussian Shading.
  • Imprinting only in the final five steps raises PSNR from 22.40 to 26.05 dB and nearly halves LPIPS.
  • Frequency patterns, bit-margin correction and alignment search add resistance to cropping, rotation and regeneration.
  • SD3 and FLUX required adapting the watermark modules; no public code yet (paper page).

Diffusion-image watermarking usually carries a computational or operational cost. Initial-noise schemes preserve the generator’s weights but require dozens of reverse diffusion steps for detection. Weight-based schemes decode quickly but require training for each model checkpoint. DiffMark, developed by Hong-Hanh Nguyen-Le and colleagues, injects arbitrary 256-bit keys during the final denoising steps of a frozen generator and reads them without reversing the diffusion process.

The revised paper reports fast extraction and high detection rates across Stable Diffusion 2.1, Stable Diffusion 3, and FLUX. It was accepted at the Second Workshop on Technical AI Governance Research at ICML.

Property Reported result
Payload 256 bits per image
Extraction time 89.4 ms
Detection 99.4% to 100% average TPR at 1% FPR across tested distortion, removal, and adversarial attacks
Generators SD 2.1, SD3, and FLUX, with architecture-specific watermark modules
Base generator Frozen during watermark training and use

Why verification can cost a generation

Diffusion generators begin with random noise and repeatedly denoise it into an image. Tree-Ring places a pattern in the Fourier representation of that initial noise. It is a zero-bit scheme, meaning the detector reports whether a watermark is present but carries no message. Gaussian Shading embeds a recoverable message in the starting noise.

Both approaches reconstruct the initial noise from the finished image during verification. DDIM inversion runs the sampling process in reverse, commonly for about 50 steps, so checking an image can cost about as much computation as generating one. That expense constrains services scanning large upload volumes.

Fine-tuning the generator offers faster detection because every output carries a learned signature. A checkpoint, the stored set of model weights, remains tied to the watermark training process. Supporting another architecture or checkpoint can require retraining, and assigning a distinct key to every image becomes harder.

Five late steps carry the key

DiffMark’s encoder maps a 256-bit key to a perturbation in the generator’s latent space, the compact representation processed during diffusion. The sampler adds that perturbation near the end of denoising, when the image structure is already established. A learned decoder then recovers the bits from the finished image through its VAE representation.

Pro article

This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

Trending
  • No trending articles

Comments

avatar

Next Reads