US 11,836,643 B2
System for secure federated learning
Kumar Sharad, Heidelberg (DE); Ghassan Karame, Heidelberg (DE); and Giorgia Azzurra Marson, Heidelberg (DE)
Assigned to NEC CORPORATION, Tokyo (JP)
Filed by NEC Laboratories Europe GmbH, Heidelberg (DE)
Filed on Mar. 8, 2019, as Appl. No. 16/296,380.
Prior Publication US 2020/0285980 A1, Sep. 10, 2020
Int. Cl. G06N 3/098 (2023.01); G06N 5/043 (2023.01); G06N 20/20 (2019.01); G06F 21/60 (2013.01)
CPC G06N 5/043 (2013.01) [G06F 21/606 (2013.01); G06N 3/098 (2023.01); G06N 20/20 (2019.01)] 11 Claims
OG exemplary drawing
 
1. A method for performing federated learning, comprising:
initializing, by a server, a global model G0;
sharing, by the server, G0 with a plurality of participants (N) using a secure communications channel;
selecting, by the server, n out of N participants, according to filtering criteria, to contribute training for a round r;
partitioning, by the server, the selected participants n into s groups;
informing, by the server, each participant about the other participants belonging to the same group;
obtaining, by the server, aggregated group updates AU1, . . . , AUg from each group for the round r, wherein, each participant locally uses local training data to generate a local update Ui needed to obtain a local model Li from the previous global model Gr-1 and wherein each participant belonging to the same group k=1, . . . , g, executes a secure aggregation protocol to combine all of their updates U1k, . . . , Usk, to obtain the aggregated group update AUg, wherein the local update Ui includes the size di of the local training data;
comparing, by the server, the aggregated group updates and identifying suspicious aggregated group updates probabilistically using machine learning for the round r;
combining, by the server, the aggregated group updates by assigning a weight to each group update AU1, . . . , AUg, averaging the weighted group updates to obtain an aggregated update Ufinal and adding the aggregated update to a previous parameter vector, wherein the weight assigned to suspicious aggregated group updates is lower than the weight assigned to group updates that are not suspicious;
deriving, by the server, a new global model Gr from the previous model Gr-1 and the aggregated update Ufinal; and
sharing, by the server, Gr with the plurality of participants.