A privacy-preserving revocable framework in the deep-learning-as-a-service platform system based on non software as a service
Ganglin Zhang, Yongjian Liao, Shijie Zhou · 2021 IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech) · 2021
With the gradual maturity of deep learning, it is possible to make deep learning models as a service on the cloud. However, privacy issues are the key problem of making deep learning models as cloud services. According to the service mode of cloud computing, the service mode of deep learning is divided into software-as-a-service-based systems and non-software-as-a -service- based systems. In non -software-as-a-service-based systems, the cloud computing provider can hold the deep learning model, but the deep learning model can be revoked by the model owner. Many solutions have been proposed to protect privacy in a cloud server that makes deep learning models as cloud services. However, there is no study on how to revoke the cloud server's authority to use the deep learning model. To solve the privacy issues and support the revocability of the model in non-software-as-a-service-based systems, we propose a privacy-preserving revocable framework based on fully homomorphic encryption and digital signature. This framework not only protects the privacy of input data, output results, and model parameters but also allows the model owner to revoke the cloud's right to use the model. Then, the framework is proved to conform to our security definition through security analysis. Next, an instance based on the existing fully homomorphic encryption scheme and the digital signature scheme is constructed. At last, experiments are used to prove that the framework is feasible in practice and estimate the computing time cost of the framework.