Testing Attacks on Structural Text Watermarking Techniques

Every day more and more content is published on the internet in social media sites, cloud sharing services or websites and this content becomes easy pickings for people who want to steal it and make it theirs. When this sort of “content theft” occurs it is very difficult to trace the stolen piece of content back to its original author. Another linked issue is that publishers and content creators are interested in a lightweight and easy to use solution to this problem in order to protect their works against different types of attack. The most common solution to this problem is Digital Watermarking, defined as the process of digitally embedding a certain amount of information (the watermark) into a piece of multimedia content such as a picture, video or written document. The most apparent form of Digital Watermark is a photographer’s name printed on top of their picture in order to protect it from improper use, as it would carry the watermark if it were to be copied and pasted in another website, thus linking the picture to its owner. Different pieces of media need different watermarking techniques because, for example, an image has different properties from a piece of text. Speaking of the latter, text watermarking is especially difficult, because text as a medium has a low embedding rate for external information: for example if one inserts a new word into a piece of text, any attacker would be able to understand where the watermark is and how to delete it. This means that text watermarking is done through other more discreet means like converting text to an image and watermarking the result, using synonyms or with the help of structural watermarking methods. This thesis will focus on text watermarking because of its abundance in literature and the inherent omnipresence of text on the internet which make it almost omnipresent in blogs, social networks, forums and scientific journal’s websites.