Is the identification of AI models really that reliable? Most fingerprint recognition experiments are based on an assumption - that the model hosting party is benign and will not actively remove watermarks or identification marks. Sounds quite idealistic.

But what is the reality? In an ecosystem where models are traded, merged, forked, and repackaged, this assumption simply does not hold. Once a model enters the circulation stage, the risk of the identification being tampered with, removed, or even forged sharply increases. Your identification mechanism may perform perfectly in the lab, but in real-world scenarios, it becomes a mere decoration. This is also why model security requires deeper technical design – it cannot rely solely on good-faith assumptions.