Salvato in:
Dettagli Bibliografici
Autore principale: Wang, Shuchan
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:https://arxiv.org/abs/2605.16239
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866918504586805248
author Wang, Shuchan
author_facet Wang, Shuchan
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2605_16239
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Dynamics-Level Watermarking of Flow Matching Models with Random Codes
Wang, Shuchan
Machine Learning
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.
title Dynamics-Level Watermarking of Flow Matching Models with Random Codes
topic Machine Learning
url https://arxiv.org/abs/2605.16239