Back-propagation is an effective method for neural network learning. To improve the accuracy of the learning result, in practice multiple parties may want to collaborate by jointly executing the back-propagation algorithm on the union of their respective data sets. During this process no party wants to disclose her/his private data to others for privacy concerns. Existing schemes supporting this kind of collaborative learning just partially solve the problem by limiting the way of data partition or considering only two parties. There still lacks a solution for more general and practical settings wherein two or more parties, each with an arbitrarily partitioned data set, collaboratively conduct learning.
In this paper, by utilizing the power of cloud computing, we solve this open problem with our proposed privacy preserving back-propagation algorithm, which is tailored for the setting of multiparty and arbitrarily partitioned data. In our proposed scheme, each party encrypts his/her private data locally and uploads the ciphertexts into the cloud. The cloud then executes most of the operations pertaining to the learning algorithms with ciphertexts but learns nothing about the original private data. By securely offloading the expensive operations to the cloud, we keep the local computation and communication costs on each party minimal and independent to the number of participants. To support flexible operations over ciphertexts, we adopt and tailor the BGN ‘doubly homomorphic’ encryption algorithm for the multiparty setting. Thorough analysis shows that our proposed scheme is secure, efficient and scalable.