New Generative Model Offers Robust Time-Series Watermarking.

Watermarking is an essential technique for establishing provenance and authenticity in generative models, but its application to multivariate time series data has been problematic. Existing watermarking detectors, which often rely on globally coupled re-encoding, are vulnerable to post-editing attacks, leading to unreliable detection due to bidirectional drift in the null distribution. This instability can cause non-watermarked samples to be incorrectly flagged or genuine watermarks to be missed. This paper proposes that reliable detection requires each recovered unit to depend solely on a limited temporal neighborhood, rather than the entire sequence. Guided by this principle, the researchers developed L-VQVAE, a generative model where discrete tokens are produced from short, contiguous windows of time. Building on this, they introduce LVQMark, a watermarking method that combines logit-bias injection with a robust re-encoding mechanism for detection at the time of attack. Experiments conducted across four diverse benchmarks—including finance, energy, and neuroimaging data—demonstrate the effectiveness of their approach. LVQMark not only preserves the quality of generated time series but also significantly stabilizes both detection power and false-positive behavior, even when subjected to various post-editing attacks. This marks a crucial step towards reliable provenance tracking for generative time-series models.