The rise of Machine Learning (ML) has opened new business opportunities, particularly through Machine Learning as a Service (MLaaS), where costly models like Deep Neural Networks (DNNs) can be outsourced. However, this also raises concerns about model piracy. To protect against unauthorized use, watermarking techniques have been developed. One such method, null embedding by Li et al., disables the model if pirated but reduces classification accuracy. This paper proposes modifications to the null-embedding technique that reduce this impact and keep the classification accuracy close to that of a non-watermarked model.

In this work, we focus on watermarking time series datasets and explore one of the techniques known from audio-watermarking, namely Discrete Wavelet Transform (DWT) based watermarking, to investigate its effectiveness. We adapt (Attari and A. Shirazi, 2018) and embed a bit stream into a time series dataset by calculating the DWT coefficients and modifying their magnitudes for embedding. Our experimental results on two real-world datasets show good robustness against a small range of data modification attacks but lack capability in larger-scale attacks. We believe that our work could initiate a new research direction on dataset watermarking using well-known techniques from signal processing.