I suppose the approach is to create a machine learning dataset that maps hi-res scans of sample documents to the printers that produced them. If the resulting classifier can accurately id the printer, you have probably found a watermark, but it might just be natural variations in the manufacturing.
The difficulty in this approach is that you have an extremely large number of classes. Each printer is its own class. Typically, as the number of classes goes up, accuracy goes done. That isn't to say it isn't possible, but it would require a lot of custom hacks to any learning algorithm.
Also to convince anyone that it works, you would need to test it out on an extremely large number of printers, including ones of the same model. In practice that could be expensive.
