Multi-Party Computation on Machine Learning

Summary: A Trail of Bits intern built a three-party multi-party computation tool implementing perceptron and support vector machine training across private datasets, using secret sharing over integers modulo a prime and Beaver Triples for secure multiplication. One iteration of each algorithm requires around 15 gates, compared with thousands of addition and multiplication gates when synthesizing AES, and the tool achieved arbitrary fixed-point precision with classification accuracy matching the raw algorithms.
What you should know: The pitch is that three organizations can train a model on data none of them will hand over, and the barrier is no longer circuit size: one training iteration fits in roughly 15 gates. The cost that remains is communication, since secure multiplication requires the parties to interact and does not parallelize well.

Read the original at Trail of Bits Blog ↗