6533b839fe1ef96bd12a630b

RESEARCH PRODUCT

Automatic Identification of Watermarks and Watermarking Robustness Using Machine Learning Techniques

subject

description

The goal of this article is to propose a framework for automatic identification of watermarks from modified host images. The framework can be used with any watermark embedding/extraction system and is based on models built using machine learning (ML) techniques. Any supervised ML approach can be theoretically chosen. An important part of our framework consists in building a stand-alone module, independent of the watermarking system, for generating two types of watermarks datasets. The first type of datasets, that we will name artificially datasets, is generated from the original images by adding noise with an imposed maximum level of noise. The second type contains altered watermarked images obtained from the original ones by using different transformations. The module also performs an automatic labeling process of these data, building watermarks’ containers. Then, many models can be built using the watermarks containers and different ML techniques. Comparing the performances of all the obtained models allows the choice of the best model, or provides details for building ensemble learning. To validate the proposed framework, we conducted experiments using a particular watermarking system, built by us and many models based on artificial neural networks (ANN) and support vector machines (SVM). As a side result we identified a possible methodology for evaluating the robustness of a watermarking system, by using ANN and the two types of datasets generated in our proposed methodology.