Dynamics-Level Watermarking of Flow Matching Models with Random Codes

TL;DR

This paper proposes a novel watermarking method for flow matching generative models by embedding watermarks into their continuous dynamics, ensuring reliable detection without affecting output quality.

Contribution

It introduces a dynamics-level watermarking technique using random coding in the velocity field of flow models, enabling black-box detection without altering generated data.

Findings

Reliable message recovery demonstrated on MNIST and CIFAR-10

Watermarking preserves generation quality

Decoding accuracy is at chance level without the secret key

Abstract

We introduce a dynamics-level approach to watermarking generative models. Rather than embedding signals into model weights or outputs, we embed the watermark directly into the learned continuous dynamics -- the velocity field of a flow matching model. We formulate this as random coding over a continuous channel: a key-dependent perturbation is added during training, and the message is recovered at detection time from black-box queries. The perturbation is designed to leave the generated distribution unchanged. Experiments on MNIST and CIFAR-10 across different architectures confirm reliable message recovery, preserved generation quality, and chance-level decoding accuracy without the secret key.