| Recently,with the development of big data and cloud computing technology,artificial intelligence(AI)technology has penetrated all aspects of transportation.As a product of the deep integration of the transportation industry with the new generation of information technologies such as artificial intelligence,big data,and the Internet of Things,autonomous driving provides a new solution for traffic safety,road congestion,energy consumption,and environmental pollution,and is expected to become the most valuable application scenario in the AI boom.The rapid development of autonomous driving technology and increasingly complex on-board systems pose new challenges to system safety.On the one hand,the massive information collected by onboard sensors provides strong support for intelligent upgrading of the automatic driving system.However,the traditional public key cryptosystem based on computational complexity cannot meet the requirements of secure transmission and real-time response of massive image information currently collected.Compared with plaintext,the traditional encryption method needs more time and resources,violating the real-time response of automatic driving system.On the other hand,resource-intensive AI algorithms usually need the computing and storage resources provided by the cloud platform to complete the training and prediction of AI models.But the infrastructure used by remote services,including cloud platforms,is usually managed by an untrusted third party.How to safely integrate AI into autopilot systems has become an urgent problem to be solved in its development process.Aiming at the above problems,this paper focuses on the analysis of large-scale image security transmission technology and trusted AI algorithms for automatic driving.Research is carried out from the following three aspects:(1)Aiming at the problem that the traditional data protection method based on public key cannot meet the real-time and safe transmission of massive data in an automatic driving system.This paper presents a secure and efficient image-sharing scheme based on visual cryptography(VC),which solves the contradiction between automatic driving data protection and real-time response.By combining the threshold characteristics of secret sharing with the visual friendly characteristics of images,visual cryptography transforms complex encryption and decryption operations into simple Boolean operations,which provides a key-free solution for the privacy protection of massive images,eliminates the performance loss caused by traditional encryption methods,and provides a new theoretical method for data protection of big data applications such as autonomous driving.(2)To solve the problem of noise interference in encrypted images,this paper uses the transfer learning method to transfer the high-precision model parameters trained from large-scale data sets to the recognition model for lossy data such as encrypted images.The strong generalization ability of transfer learning eliminates the problem that the recognition accuracy decreases due to the interference signal in VC,and provides a new technical scheme for the high-precision recognition of encrypted data sets.(3)Aiming at the problem that the AI algorithm running in the untrusted cloud platform is vulnerable to attack.This paper proposes a runtime protection scheme for AI algorithms based on Trusted Execution Environment(TEE).Using dynamic dependency analysis and Lib OS(Library Operating System)technology,the existing AI recognition system can run directly in the safe area without code refactoring,to prevent sensitive data from being sniffed in the running process.This scheme shields the complexity of the traditional privacy protection technology and creates a more friendly running model for the application in TEE.In this paper,the remote authentication mechanism is used to further prevent the data of the shared recognition template from being stolen by the untrusted third party,including the operating system,to protect the security and credibility of the AI recognition system running on the untrusted cloud platform.The extensive experiments on traffic datasets show that the scheme proposed in this paper not only protects the secure transmission and reliable operation of largescale image data,but also maintains the high-precision recognition characteristics of the AI model. |
