The rapid development of Deep Neural Networks (DNNs) has propelled Machine Learning as a Service (MLaaS) into a viable business model. However, during deployment in cloud or local environments, DNNs are often exposed to untrusted environments, facing security risks such as theft, resale, misuse, and copyright infringement. Unfortunately, existing mainstream defense mechanisms suffer from limitations in flexibility and applicability, failing to provide comprehensive protection for model functional services. To address these challenges, this article leverages the inherent capability of backdoor attacks to enforce specific model decisions, proposing a novel model protection framework based on input-sensitive neural networks. This sensitivity means that the model only exhibits semantic prediction capabilities for legitimate images embedded with copyright-related triggers, while treating all other inputs as illegitimate images and outputting semantically irrelevant prediction labels. The security of the framework relies entirely on the key used in the trigger generation process, adhering to Kerckhoff’s principle. Extensive experiments demonstrate that the proposed framework satisfies practical requirements in terms of performance fidelity, scenario universality, computational efficiency, and security robustness.