A Practical Scheme for Two-Party Private Linear Least Squares
This work addresses privacy-preserving machine learning for distributed sensitive data, but it is incremental as it focuses on a specific subclass of algorithms.
The paper tackles the problem of training a linear least squares model across two parties without a trusted third party, using gradient descent with additive homomorphic encryption, resulting in a protocol requiring two communication rounds per gradient descent step.
Privacy-preserving machine learning is learning from sensitive datasets that are typically distributed across multiple data owners. Private machine learning is a remarkable challenge in a large number of realistic scenarios where no trusted third party can play the role of a mediator. The strong decentralization aspect of these scenarios requires tools from cryptography as well as from distributed systems communities. In this paper, we present a practical scheme that is suitable for a subclass of machine learning algorithms and investigate the possibility of conducting future research. We present a scheme to learn a linear least squares model across two parties using a gradient descent approach and additive homomorphic encryption. The protocol requires two rounds of communication per step of gradient descent. We detail our approach including a fixed point encoding scheme, and one time random pads for hiding intermediate results.