US 12,244,707 B2
System and method for sharing an incrementally trained machine learning (ML) model from an edge device to one or more other edge devices in a peer to peer network
Subash Sundaresan, Fremont, CA (US)
Assigned to Subash Sundaresan, Fremont, CA (US)
Filed by swarmin.ai, Fremont, CA (US)
Filed on Nov. 19, 2020, as Appl. No. 16/953,095.
Claims priority of provisional application 62/972,580, filed on Feb. 10, 2020.
Prior Publication US 2021/0250166 A1, Aug. 12, 2021
Int. Cl. G06N 3/098 (2023.01); G06N 3/096 (2023.01); G06N 20/00 (2019.01); H04L 9/08 (2006.01); H04L 9/30 (2006.01); H04L 67/1074 (2022.01)
CPC H04L 9/088 (2013.01) [G06N 3/096 (2023.01); G06N 3/098 (2023.01); G06N 20/00 (2019.01); H04L 9/30 (2013.01); H04L 67/1074 (2013.01)] 12 Claims
OG exemplary drawing
 
1. A processor-implemented method for sharing an incrementally trained machine learning (ML) model from a first edge device to other edge devices in a peer to peer network, said method comprising:
generating a unique persistent file of the incrementally trained ML model at the first edge device by removing one or more details associated with specific event data items used to incrementally train a base version of a first ML model, wherein the one or more details comprises details associated with the user, wherein the first ML model comprises at least one parameter and at least one weight associated with the at least one parameter, wherein the unique persistent file optimizes a payload by including the at least one parameter and the weight associated with the at least one parameter that has changed beyond a predetermined configurable threshold;
creating a hashed token for each event data item used to incrementally train the ML model, wherein the hashed token comprises at least (i) a unique device ID of the first edge device, (ii) a timestamp of the event, and (iii) an event file size, and wherein the hashed token is stored at the first edge device for comparison against subsequently submitted event data items to prevent re-use of previously submitted event data during incremental training at the same or different edge devices;
encapsulating the unique persistent file with a unique metadata to ensure that event data used for the incremental training of the first ML model at the edge device is not re-used at the same or different edge devices, wherein the metadata comprises the hashed token comprising at least (i) the unique device ID of the first edge device, (ii) the timestamp of the event, and (iii) the event file size;
enabling an authentication of the incrementally trained ML model generated at the first edge device by encrypting the unique persistent file of the incrementally trained ML model, wherein the encryption is performed using a key corresponding to the first edge device to prevent injection of unauthorized versions of the ML model into the peer-to-peer network; and
incrementally training a second ML model at other edge devices in the peer-to-peer network by transmitting the incrementally trained ML model in the encrypted unique persistent file from the first edge device to the other edge devices in the peer-to-peer network.