The increasing need for privacy in machine learning has driven interest in cryptographic methods that enable computation on sensitive data. Secure multi-party computation and fully-homomorphic encryption are two prominent approaches, but their practical trade-offs are often unclear for real-world use. This work presents a pragmatic benchmark and comparison of these technologies for machine learning inference. Using representative open source frameworks, we evaluate performance across basic operations, distance metrics, regression models, and common deep learning architectures. The evaluation accounts for realistic deployment conditions, including communication overhead and partial plaintext execution. The results show that secure multi-party computation scales well to large and complex models but is strongly affected by network latency, while fully-homomorphic encryption is easier to deploy and performs well for small models and regression tasks. We conclude that the two approaches are complementary rather than competing and provide concrete guidance to help practitioners select the most suitable technology for privacy preserving machine learning in practice.
Taubert et al. (Tue,) studied this question.