L-VQVAE 与 LVQMark：面向时间序列水印的局部 token 化生成模型
A Locally Tokenized Generative Model for Robust Time-Series Watermarking
AI 导读
研究提出 L-VQVAE 生成模型与 LVQMark 水印方法，将时间序列编码为基于短连续窗口产生的离散 token，结合 logit-bias 注入与稳健重编码进行攻击时检测。在覆盖金融、能源和神经影像的四个基准上，该方法在保持生成质量的同时，使检测能力与假阳性表现对后编辑攻击更稳定。相关工作入选 NeurIPS 2026。
正文 · 原文
Abstract:Watermarking is a central tool for provenance in generative models, yet its application to multivariate time series remains hindered by reliability failures under post-editing attacks. We show that existing detectors, which rely on globally coupled re-encoding, suffer from bidirectional drift of the null distribution: post-editing attacks can shift the $z$-score of non-watermarked samples in either direction, invalidating clean-calibrated thresholds. We argue that this instability is a property of the re-encoding, and that reliable detection requires each recovered unit to depend only on a bounded temporal neighborhood. Guided by this principle, we propose L-VQVAE, a generative model in which each discrete token is produced from a short contiguous window, and LVQMark, a watermarking method over this token space that combines logit-bias injection with robust re-encoding for attack-time detection. Experiments on four benchmarks spanning finance, energy, and neuroimaging show that our approach preserves generation quality while stabilizing both detection power and false-positive behavior under post-editing attacks.
| Comments: | NeurIPS 2026 |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| ACM classes: | I.2.6; K.6.5; G.3 |
| Cite as: | arXiv:2608.19727 [cs.LG] |
| (or arXiv:2608.19727v2 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2608.19727 arXiv-issued DOI via DataCite |
Submission history
From: Dongbin Kim [view email]
[v1]
Thu, 20 Aug 2026 07:25:41 UTC (538 KB)
[v2]
Thu, 5 Nov 2026 02:11:50 UTC (550 KB)
来源：arXiv cs.AI · arxiv.org