To do that you need to figure out what parts are noise (watermark), simply diffing them would just give you a different noise pattern which can still be analyzed depending on how the watermark is encoded.