EPDL: An efficient and privacy-preserving deep learning for crowdsensing

Deep learning has achieved remarkable success in the field of crowdsensing. The success of deep learning is inseparable from the amount of data. However, since the data uploaded by users involved in deep learning are usually correlated with individuals’ personal information, data owners may be reluctant to provide their data. Existing works on privacy-preserving deep learning primarily rely on fully homomorphic encryption primitives or oblivious transfer, which generates a lot of computation and communication costs to the participating entities. In this paper, we propose a non-interactive privacy-preserving deep learning scheme, named EPDL, to solve the above privacy and efficiency issues which means we can train the model more efficiently while protecting the privacy of image data. By employing a cloud platform and exploiting the homomorphic properties of an additively homomorphic cryptosystem, EPDL enables the deep learning models to be trained in an efficient and privacy-preserving manner without any data owner involved in the training process. Detailed security analysis demonstrates the privacy of data and models is safeguarded by EPDL. Extensive experiments based on real-world data sets show EPDL outperforms existing schemes whether in computation costs or communication overhead.