Researchers have developed a new method for robust time-series watermarking. The approach uses a generative model called L-VQVAE to encode each data point in a short window of its neighbors, rather than relying on global connections. This makes the model more resistant to post-editing attacks that can shift the distribution of non-watermarked samples. The authors also propose a new watermarking method called LVMark that combines logit-bias injection with robust re-encoding for attack-time detection. Experiments show that this approach preserves generation quality while improving detection power and reducing false positives. --- Why it matters: This matters to researchers in AI because it addresses a significant challenge in time-series data, which is often used in applications such as finance, energy, and healthcare. The ability to robustly watermark time-series data can help ensure the provenance of generated models and prevent tampering attacks. Source: https://arxiv.org/abs/2608.19727