Secure multiparty computation, often abbreviated MPC, is a branch of cryptography that lets several parties compute a function over their combined data without any party revealing its own input to the others. The classic illustration is a group of people who want to learn their average salary without anyone disclosing what they earn. MPC makes this possible: at the end everyone learns the agreed-upon answer, the average, and nothing else about anyone’s individual figure. The result is computed correctly even though the raw inputs were never pooled or exposed.
In machine learning, MPC is the tool that lets multiple data holders train or query a model jointly while keeping their datasets private from one another. A concrete and widely used instance is secure aggregation, the protocol Google introduced in 2017 for federated learning: many devices each hold a model update derived from their private data, and MPC techniques let a server compute the sum of all the updates while being mathematically unable to see any single device’s update. Each contribution is masked with random values designed to cancel out only when everything is added together, so individual data stays hidden but the useful aggregate emerges.
MPC complements the other privacy tools. Differential privacy limits what a published result reveals; homomorphic encryption lets one party compute on another’s encrypted data; MPC coordinates computation across several mutually distrustful parties. They are often layered together.
For a business reader, MPC is what makes genuine collaboration on sensitive data possible without a trusted intermediary. Competing banks can jointly detect fraud patterns, hospitals can pool insight from patient records, or partners can compute shared analytics, each contributing data that the others, and even the coordinator, never get to see. The cost is added complexity and computation, so it is reserved for cases where the inability to expose raw data is a hard requirement.