Secure Multiparty Computation (MPC) has attracted more and more attention nowadays. Various MPC softwares have been proposed by the research community, but two major challenges still exist when deploying them to real industry systems:
- Insufficient MPC machine learning algorithms for model training. Current MPC ML training solutions mainly focus on simple models (e.g., linear regression, logistic regression), but their accuracy is unsatisfactory in complicate applications, sometimes it’s even lower than that trained from a single party, which contradicts the target of MPC.
- Lack of user-friendly interfaces. Many academic MPC codes are purely designed for efficiency (e.g. developed in C/C++ language, hand-crafted I/O channels), which makes them unscalable, and hard to use for non-expert users. Programmers would be more comfortable with a scientific scripting language like python for machine learning.
- More powerful MPC ML model training, including but not limited to: Tree-based models (e.g., decision tree，random forest，GBDT), SVM, CNN, RNN, feature selection and extraction (e.g., hypothesis testing, subset selection, PCA), regularization methods (e.g., l1, l2).
- MPC library with user-friendly interface (e.g. python-like Domain Specific Language) for non-expert programmers.
Related Research Topics
- ABY MPC framework
- Privacy-preserving Machine Learning
- Computer language and Compilers