Secure aggregation

A cryptographic protocol that computes an aggregate value (e.g., model gradients) from multiple parties without revealing any individual input to the central server.

Secure Aggregation (SecAgg) is a core privacy primitive for Federated Learning (FL). It enables a central server to compute a collective sum, like a global model update, from numerous client-side contributions: no individual client's data is exposed. The original protocol (Bonawitz et al., 2017) uses Secure Multi-Party Computation (SMPC) to mask individual inputs, ensuring the server only decrypts the final, aggregated result. This design is robust and communication-efficient, specifically engineered to tolerate high client dropout rates—up to one-third of participants—without compromising the overall computation or privacy guarantee.

https://research.google/pubs/pub46522/