[Download postscript version]

Improvements on the schemes

  Now that we have seen all the issues arising when analyzing today's watermarking schemes, we will present possible improvements.

First we observed that watermarking is regarded as being critically important to solving the digital image copyright protection problem. But when copyrighting traditional paintings or photographs, watermarks are not used. So why should we rely so strongly on watermarks if they are complicated to handle and so easy to defeat? This forms the basis of a slightly different approach to copyright protection. A simple approach is based on a model analog to today's copyright office [Per97]. Every user registers the images to copyright at the copyright office. In case of dispute, the court of law can compare the disputed image with the registered image. So far we are not using any watermarks. But watermarks are still useful for discovering fraud on the network because of the difficulty of automatic comparison of images. Watermarks can be extracted automatically when the secret key is known. We can therefore look at watermarking as a method for matching images efficiently rather than as a security device. We have therefore shifted the importance of watermarking to fraud detection.

It is most important that watermarks used for the purpose of image search in this way are kept separate from those used for the purpose of author identification. Using the same watermark for both purposes drastically weakens the security of the scheme. For image search watermarks, we do not need exact retrieval, as we can correlate the mark against those of the image(s) we are searching for, but we do need much greater robustness against watermark detection and removal. For public- information watermarks, the need for robustness is much less, but the need to retrieve the watermark intact is much greater.

The problems addressed with web-spiders can be solved by running the spider on a large ISP proxy server. For example the AOL proxy server could forward all the images that it has seen to the spider to check. With this method, access controlled sites could be checked as well as paying sites. Mallory could not refuse to send the image to the proxy server because it would affect a large number of customers that could not receive the image.

The problem of proving fraud to a court of law could be solved if each web-server would sign every response with his private key. We would therefore have a proof that the server sent the message to us. A payment protocol, such as Netbill, would also take care of this.

We stated earlier that the UI should protect the user from handling the watermarks erroneously. One possible improvement is to include a watermark strength indicator on the screen, so the user sees how the watermark reacts to image transformations. Ideally the watermark embedding is the last step before publishing the image. Therefore the software could delay the embedding until the user wants to save the image. Equally when loading the image, the watermark could be extracted and the user would always work with an unwatermarked image. The watermarking would therefore become a transparent operation to the user.