Highly Scalable Beaver Triple Generator from Additive-only Homomorphic Encryption
In a convolution neural network, a composition of linear scalar product, non-linear activation function and maximum pooling computations are intensively invoked. As such, to design and implement privacy-preserving, high efficiency machine learning mechanisms, one highly demands a practical crypto tool for secure arithmetic computations. SPDZ, an interesting framework of secure multi-party computations is a promising technique deployed for industry-scale machine learning development if one is able to generate Beaver (multiplication) triple offline efficiently. This paper studies secure yet efficient Beaver triple generators leveraging privacy-preserving scalar product protocols which in turn can be constructed from additive-only homomorphic encryptions(AHEs). Different from the state-of-the-art solutions, where a party first splits her private input into a shared vector and then invokes an AHE to compute scalar product of the shared vectors managed by individual MPC server, we formalize Beaver triple generators in the context of 2-party shared scalar product protocol and then dispense the generated shares to MPC servers. As such, the protocol presented in this paper can be viewed as a dual construction of the state-of-the-art AHE based solutions. Furthermore, instead of applying the Paillier encryption as a basis of our previous constructions or inheriting from somewhat homomorphic encryptions, we propose an alternative construction of AHE from polynomial ring learning with error (RLWE) which results in an efficient implementation of Beaver triple generators.