Secure aggregation is a cryptographic protocol used in federated learning and similar distributed systems. It lets a server combine many client updates into one result without being able to read each participant’s individual update. The server sees the aggregate, but not the raw contribution from any single device, tenant, or user.
It matters because it reduces exposure of sensitive training signals in transit and at the aggregation point. Defenders use it to limit what an operator, cloud service, or compromised coordinator can learn from model updates. It is common in privacy-preserving machine learning, where clients send encrypted or masked values that cancel out only when combined. But secure aggregation is not a full privacy guarantee: it does not prove that the cohort was safe, that the updates were honest, or that the final model meets a privacy budget. Attackers can still exploit weak participant selection, small cohorts, or metadata leaks around the aggregation process.