| Nowadays,machine learning technology is developing rapidly,and at the same time,the scenarios of sharing machine learning models are increasing.On the one hand,users are reluctant to share data because of privacy and other reasons to form data silos,while model sharing can realize a new transaction paradigm in which the original data is not out of the library and the data is available but not visible,breaking the data silos among users.On the other hand,because the development of machine learning is also leaning toward more complex computation,machine learning algorithms now require increasingly high-performance devices for processing,and users are becoming hardpressed to afford the escalating costs alone.Thus sharing machine learning models has become an important way to popularize,make widely available,and apply contemporary AI.However,shared machine learning models pose certain security issues,and malicious code may exist in the models,generating attacks that can lead to device compromise and thus cause serious cybersecurity incidents.Therefore,this thesis investigates the security vulnerabilities of machine learning models in sharing scenarios.In this thesis,by studying and deeply researching the shared model process,we find that there are security vulnerabilities in the models under sharing,and machine learning models containing malicious codes produce behaviors beyond expectations and jeopardize user security.Therefore this thesis details the way to perform malicious code attacks in model sharing and also proposes a method to automate the generation of malicious models.For the existing security problems,on the one hand,this thesis designs a signature based pre detection method and a Hook based execution detection method,which can achieve rapid detection and accurate blocking of malicious models.On the other hand,this thesis also proposes an intelligent detection method based on machine learning,which is capable of mining representations and generating reliable feature sets to detect unknown threats.The multiple methods complement each other to further ensure user security during model sharing.The experiments show that the method of generating malicious models proposed in this thesis is effective,it has good invisibility and can threaten the attacked user.In terms of effectiveness,the experiments show that the pre detection and execution detection methods can accurately block attacks before the malicious machine learning model generates abnormal behavior.While the machine learning-based detection is compared with existing methods in various aspects,and also achieves a good overall performance. |
