TimeWak : Temporal Chained-Hashing Watermark for Time Series Data

Zhi Wen Soi, Chaoyi Zhu, Fouad Abiad, Aditya Shankar, Jeroen M. Galjaard, Huijuan Wang, Lydia Y. Chen · 2025

Read TimeWak : Temporal Chained-Hashing Watermark for Time Series Data by Zhi Wen Soi, Chaoyi Zhu, Fouad Abiad, Aditya Shankar, Jeroen M. Galjaard, Huijuan Wang, Lydia Y. Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-06407.txt.

Introduction

Soi and co-authors watermark synthetic time series so their source can be proven later, a need that grows as fake health records circulate. Diffusion models now mint realistic series directly in data space, where older latent space watermarks cannot reach. Their 2025 method, TimeWak, chains hashes along the time axis and stamps the mark inside the data itself.

A near exact inversion trick handles the uneven reconstruction errors that detection must survive across features of different scales. The authors bound that inversion error for multi feature series while keeping the mark detectable. On five datasets, the method improved realism scores by about sixty two percent against the strongest rival and stayed readable after editing attacks. Data quality and traceability, they argue, can coexist.

Catalog record

CREATORZhi Wen Soi, Chaoyi Zhu, Fouad Abiad, Aditya Shankar, Jeroen M. Galjaard, Huijuan Wang, Lydia Y. Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE99 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-06407

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TimeWak : Temporal Chained-Hashing Watermark for Time Series Data Zhi Wen Soi, Chaoyi Zhu, Fouad Abiad, Aditya Shankar, Jeroen M. Galjaard, Huijuan Wang, Lydia Y. Chen Abstract Synthetic time series generated by diffusion models enable sharing privacy-sensitive datasets, such as patients’ functional MRI records. Key criteria for synthetic data include high data utility and traceability to verify the data source. Recent watermarking methods embed in homogeneous latent spaces, but state-of-the-art time series generators operate in real space—making latent-based watermarking incompatible. This creates the challenge of watermarking directly in real space while handling feature heterogeneity and temporal dependencies. We propose TimeWak , the first watermarking algorithm for multivariate time series diffusion models. To handle temporal dependence and spatial heterogeneity, TimeWak embeds a temporal chained-hashing watermark directly within the real temporal-feature space. The other unique feature is the ϵ \epsilon -exact inversion, which addresses the non-uniform reconstruction error distribution across features from inverting the diffusion process to detect watermarks. We derive the error bound of inverting multivariate time series and further maintain high watermark detectability. We extensively evaluate TimeWak on its impact on synthetic data quality, watermark detectability, and robustness under various post-editing attacks, against 5 datasets and baselines of different temporal lengths. Our results show that TimeWak achieves improvements of 61.96% in context-FID score, and 8.44% in correlational scores against strongest state-of-the-art baseline, while remaining consistently detectable. Our code can be accessed at the repository linked below: https://github.com/soizhiwen/TimeWak . 1 Introduction Multivariate time series data drive key applications in healthcare [ 18 ] , finance [ 12 ] , and science [ 24 ] . However, access to real-world datasets is often restricted by privacy regulations, limited availability, and high acquisition costs. To address these issues, synthetic time series generated by models are increasingly adopted as practical alternatives [ 8 , 24 ] . Among generative techniques, diffusion models [ 9 ] have gained prominence for producing high-quality samples—frequently outperforming the mainstream Generative Adversarial Networks and Variational Autoencoders [ 5 , 11 ] . Beyond generation quality, traceability is equally critical—ensuring verifiability to safeguard against misuse [ 17 , 40 ] . In this context, watermarking has become the de-facto approach for tracking and auditing synthetic data [ 30 , 17 ] . The challenge lies in striking a delicate balance: embedding imperceptible signals to retain the quality of generated content whilst remaining detectable—even under post-processing [ 39 ] . Recent works embed watermarks during generation by adding them into the latent space, offerring advantages such as generality and lightweight computation [ 30 , 40 ] . However, latent-space watermarks are not always viable, especially since many state-of-the-art (SotA) time series generators operate within the real space [ 25 , 1 , 32 ] . Additionally, latent generators introduce a tradeoff in detectability—diffusion inversion and the encode–decode cycle are inherently lossy, degrading the watermarks [ 20 ] . While sampling-time watermarks have proven effective for images [ 30 ] and tables [ 40 ] , their applicability to time series data remains unexplored. Multivariate time series data possess temporal dependencies and heterogeneous features, like gender versus income. The ensuing challenges are twofold: (i) embedding watermarks directly in the real space while preserving inter- and intra-variate temporal dependencies of the generated time series, and (ii) ensuring accurate watermark detection despite the lossy nature of diffusion inversion, whilst handling mixed feature types. We propose TimeWak , the first sampling-time watermark for multivariate time series diffusion models, featuring temporal chained-hashing with \epsilon -exact inversion . TimeWak first embeds cyclic watermark patterns, i.e., the positional seeds of Gaussian noises, along the temporal direction. First, we self-hash the seeds along the temporal axis, then shuffle the seeds across features to maintain temporal correlations while preserving the unique characteristics of each feature. To ensure reliable watermark detection, we introduce an \epsilon -exact inversion strategy that makes a practical concession in the otherwise exactly invertible diffusion process: the Bi-Directional Integration Approximation (BDIA) [ 34 ] . We further provide theoretical guarantees on the resulting inversion error in Appendix C . We evaluate TimeWak against five SotA watermarking methods on five datasets under varying temporal lengths. TimeWak achieves the best detectability under six post-editing attack configurations with the minimum data quality degradation. Additionally, results show that TimeWak preserves temporal and cross-variate dependencies with high quality with near-exact watermark bit reconstruction. To summarize, we list our contributions as follows: • We propose TimeWak , the first sampling-time watermarking scheme for multivariate time series diffusion models, preserving realistic spatio-temporal dependencies while remaining detectable. • To preserve temporal characteristics and boost robustness against post-processing operations, we design a temporal chained-hashing scheme that embeds watermark seeds along the temporal direction, followed by a shuffle across features. • For robust detectability, we propose \epsilon -exact inversion by extending BDIA sampling into a real-space diffusion generator, and provide a theoretical error bound analysis. • Our extensive evaluation shows that TimeWak achieves up to 61.96% better context-FID scores and 8.44% better correlation scores compared to the strongest SotA watermarking method. 2 Related work We summarize related works on watermarking diffusion models according to their generating method (post-processing or sampling-time generation), and data modality. To the best of our knowledge, our work is the first sampling-time watermarking scheme for time series diffusion models. Watermarking diffusion models. Watermarking has become a critical solution for tracing and authenticating machine-generated content. Post-generation techniques embed watermarks after synthesis, often degrading the generated data’s quality due to direct modifications [ 4 , 37 ] . Alternatively, recent advancements embed watermarks within the training process. Studies such as Stable Signature [ 6 ] and FixedWM [ 17 ] fine-tune diffusion models to embed and extract watermarks. However, these methods modify model parameters, risking overfitting which hurts generalizability. Watermarking images. Tree-Ring (TR) [ 28 ] embeds the watermark during sampling by modifying the initial noise latent vector in the Fourier space. However, it disrupts the Gaussian noise distribution, reducing the diversity and quality of generated samples. Gaussian Shading (GS) [ 30 ] improves robustness by embedding watermarks directly in the latent space using invertible transformations. However, GS is tailored towards images and requires reversing synthetic samples into the latent space for detection—a noisy and error-prone process that limits detection accuracy. Watermarking tables. TabWak [ 40 ] watermarks tabular data in latent space using seeded self-cloning , shuffling with a secret key, and a valid-bit mechanism. However, it relies on latent models and does not account for temporal dependencies in a time series. Furthermore, its detectability is limited by the invertibility of the diffusion process and the lossy conversion to and from latent representations. 3 TimeWak Figure 1 : Overview of TimeWak . First, we assign random seeds at the beginning of each interval. \raisebox{-.9pt} {1}⃝ Temporally chained-hashing. A, B, and C ( pink ) show seeds being copied from the previous step and the feature order shuffled. \raisebox{-.9pt} {2}⃝ Shuffling the seed for each series. Positional indices are highlighted in green . \raisebox{-.9pt} {3}⃝ Constructing an initial Gaussian noise. \raisebox{-.9pt} {4}⃝ Generating multivariate time series. \raisebox{-.9pt} {5}⃝ Reversing the diffusion process. \raisebox{-.9pt} {6}⃝ Recovering the watermark seed. \raisebox{-.9pt} {7}⃝ Comparing the recovered seed with its hash. \raisebox{-.9pt} {8}⃝ Bit accuracy between the hash and

Read the complete text