AI-Finger: From Physical Unclonable Function to AI-Hardware Fingerprint

With the development of artificial intelligence (AI) and electronic technologies, large AI models are promoted to process complex tasks, e.g. natural language processing, image identification, etc. Due to resource limitations, end devices are powerless in training complex large AI models and tend to adopt AI model services provided by resource-sufficient cloud servers, named Machine Learning as a Service (MLaaS). However, in MLaaS, there exists a critical smart data leakage issue, i.e. the illegal abuse of AI models without permission. Although several existing works design authentication and protection schemes for smart data in AI models, they require permanent storage of privacy keys, which suffer from privacy key leakage and abuse issues. Moreover, existing works mainly focus on pay-per-query for MLaaS, without the ability to support pay-per-device services. To solve the above issues, we propose a physical unclonable function (PUF)-empowered AI-hardware fingerprint approach to protect AI model intellectual property. First, a PUF-empowered AI model deep protection framework is proposed, including device-specific AI-hardware fingerprint-empowered authentication and MLaaS subscription/providing. Second, we propose an AI-hardware fingerprint-enabled end-device authentication protocol to support device-bind and key-storageless authentication. Third, based on the device-bind AI-hardware fingerprint, the pay-per-device MLaaS subscription and providing scheme is designed. Experimental results verify the reliability and effectiveness of the proposed PUF-based AI-hardware fingerprint approach.