An accurate, scalable and verifiable protocol for federated differentially private averaging

Learning from data owned by several parties, as in federated learning, raises challenges regarding the privacy guarantees provided to participants and correctness of computation presence malicious parties. We tackle these context distributed averaging, an essential building block learning algorithms. Our first contribution is a scalable protocol which exchange correlated Gaussian noise along edges graph, complemented independent added each party. analyze differential our impact graph topology under colluding showing that we can nearly match utility trusted curator model even when honest party communicates with only logarithmic number other parties chosen at random. This contrast protocols local (with lower utility) or based on secure aggregation (where all pairs users need messages). second enables prove their computations without compromising efficiency protocol. construction relies standard cryptographic primitives like commitment schemes zero knowledge proofs.