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 data generation is vital for privacy-preserving data sharing, yet it necessitates robust watermarking for source verification and misuse prevention. We present TimeWak, the first generation-time watermarking algorithm designed for multivariate time series diffusion models. TimeWak utilizes a temporal chained-hashing mechanism and a mathematically bounded "$\epsilon$-exact" inversion technique to embed traceable watermarks directly in the data space. Our approach preserves temporal dynamics and data quality, achieving significant improvements in context-FID scores and robustness against common data-editing attacks.

Introduction

The rise of diffusion models has enabled the generation of high-fidelity synthetic time series data, such as patient health records and financial logs. However, tracking the provenance of this data is essential. Existing watermarking methods, typically designed for image or tabular latent spaces, struggle with time series data which often operates directly in the data space. Watermarking in this domain faces two primary challenges: spatial heterogeneity (diverse feature types) and temporal dependence (sequential correlations). Applying standard methods often destroys data utility or results in low watermark detectability due to noisy reconstruction.

Methodology: The TimeWak Pipeline

TimeWak addresses these challenges through a specialized pipeline integrated into the generation phase of the diffusion model.

Overview of the TimeWak watermark embedding and detection pipeline.

The method introduces two core innovations:

  1. Temporal Chained-Hashing: To respect the sequential nature of time series, TimeWak partitions the data into intervals and uses cryptographic hashing to chain watermark seeds along the time axis. These seeds are then shuffled across features to handle spatial heterogeneity, ensuring the watermark is deeply integrated without disrupting temporal flow.
  2. $\epsilon$-Exact Inversion: Watermark detection requires reversing the diffusion process (inversion). Standard DDIM inversion is lossy and produces non-uniform errors across different features. TimeWak employs Bi-Directional Integration Approximation (BDIA) to achieve "$\epsilon$-exact" inversion, providing a theoretical bound on reconstruction error and ensuring reliable watermark extraction.

Experimental Evaluation

TimeWak was evaluated on five diverse multivariate time series benchmarks: Stocks (Google), ETTh (Electricity transformer temperature), MuJoCo (Physics control), Energy (Appliance energy), and fMRI (Brain imaging).

Analysis of Inversion Errors

A key insight driving the design of TimeWak is the observation that inversion errors in time series diffusion models are spatially heterogeneous but temporally stable.

Reconstruction error analysis across features and time steps for Energy and MuJoCo datasets.

As shown in the error analysis, reconstruction errors vary significantly across different features (spatial heterogeneity) while remaining relatively consistent across time steps. The $\epsilon$-exact inversion specifically targets this spatial variance, which is a primary reason why traditional tabular watermarks fail in this context.

Key Results

  • Data Fidelity: TimeWak significantly outperforms baselines in data quality, showing up to a 61.96% improvement in Context-FID scores and an 8.44% improvement in correlational scores.
  • Detection Performance: The system achieves a perfect 1.0 True Positive Rate (at 0.1% FPR) with minimal samples, demonstrating high sensitivity for auditing.
  • Robustness: The watermark remains detectable under aggressive attacks, such as 30% random cropping and min-max insertions, where image-based baselines often fail.

Conclusion and Future Work

TimeWak provides a robust solution for watermarking time series diffusion models by respecting the inherent structure of the data space. While currently optimized for fixed-length windows, future work will explore extending these techniques to continuous streaming data and per-timestep watermarking for even finer granularity.

Summary updated

BibTeX
@inproceedings{soi2025timewak,
  author = {Zhi Wen Soi and Chaoyi Zhu and Fouad Abiad and Aditya Shankar and Jeroen M. Galjaard and Huijuan Wang and Lydia Y. Chen},
  title = {{TimeWak: Temporal Chained-Hashing Watermark for Time Series Data}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=akhhwQh6UV},
}
Figure