Privacy of federated QR decomposition using additive secure multiparty computation

Federated learning (FL) is a privacy-aware data mining strategy keeping the private on owners’ machine and thereby confidential. The clients compute local models send them to an aggregator which computes global model. In hybrid FL, parameters are additionally masked using secure aggregation, such that only aggregated statistics become available in clear text, not client specific updates. this context, we investigate leakage of three popular algorithms for QR decomposition, Gram-Schmidt orthonormalization, Householder algorithm Givens rotation. We show that, even when additive SMPC, rotation matrix leak raw therefore suited computation paradigm. orthonormalization relies inner vector products does points.