Print Email Facebook Twitter Image watermarking for Machine Learning datasets Title Image watermarking for Machine Learning datasets: Using SVD based image watermarking techniques to watermark numerical ML datasets Author Maesen, Palle (TU Delft Electrical Engineering, Mathematics and Computer Science) Contributor Erkin, Z. (mentor) Isler, Devris (mentor) Kellnhofer, P. (graduation committee) Degree granting institution Delft University of Technology Programme Computer Science and Engineering Project CSE3000 Research Project Date 2023-02-03 Abstract The media watermarking technique domain has had the last 30 years to develop itself. The non-media side, however, is a way newer sub-domain. [1] The data-gathering process for machine learning algorithms is a tedious and time consuming task. This becomes worse as the scale of these algorithms increases. Thus, protecting the datasets against illegal use or sale and proving they are intellectual property is useful. In this paper, we answer the question: How can image watermarking techniques be applied to classification algorithm datasets, without degrading the dataset's quality? Algorithms that use the Singular Value Decomposition (SVD) of the data are often the basis of other matrix decomposition based Image watermarking techniques. Thus if an SVD based algorithm can be applied to a machine learning dataset then the other matrix decomposition based algorithms can also be applied. This implies that a large part of the much older media targeted watermarking techniques can be applied to the non-media datasets. In this paper we apply the watermarking technique described in [2] to a machine learning dataset. This watermark provides decent imperceptibility and robustness against update, zero-out and insertion attacks but it's held back by its lackluster robustness against deletion attacks. That said, we proved that when an image watermark is found that is impervious against deletion attacks, it can be applied to the machine learning datasets. Subject Singular Value DecompositionwatermarkingMachine Learning To reference this document use: http://resolver.tudelft.nl/uuid:0fd384aa-c0c1-42bb-ac7d-b8a091a02c33 Part of collection Student theses Document type bachelor thesis Rights © 2023 Palle Maesen Files PDF Research_Paper_Palle_5007097.pdf 758.43 KB Close viewer /islandora/object/uuid:0fd384aa-c0c1-42bb-ac7d-b8a091a02c33/datastream/OBJ/view