US 12,314,864 B2
System and method for mimicking a neural network without access to the original training dataset or the target model
Eli David, Tel Aviv (IL)
Assigned to Nano Dimension Technologies, Ltd., Ness Ziona (IL)
Filed by NANO DIMENSION TECHNOLOGIES, LTD., Ness Ziona (IL)
Filed on Feb. 15, 2024, as Appl. No. 18/442,718.
Application 18/442,718 is a continuation of application No. 16/910,744, filed on Jun. 24, 2020, granted, now 11,907,854.
Application 16/910,744 is a continuation in part of application No. 16/211,994, filed on Dec. 6, 2018, granted, now 10,699,194, issued on Jun. 30, 2020.
Application 16/910,744 is a continuation in part of application No. PCT/IL2018/051345, filed on Dec. 10, 2018.
Application PCT/IL2018/051345 is a continuation of application No. 16/211,994, filed on Dec. 6, 2018, granted, now 10,699,194, issued on Jun. 30, 2020.
Claims priority of provisional application 62/679,115, filed on Jun. 1, 2018.
Prior Publication US 2024/0232635 A1, Jul. 11, 2024
Int. Cl. G06N 3/086 (2023.01); G06N 3/04 (2023.01); G06N 3/084 (2023.01)
CPC G06N 3/086 (2013.01) [G06N 3/04 (2013.01); G06N 3/084 (2013.01)] 28 Claims
OG exemplary drawing
 
1. A method to mimic a pre-trained target model at a device without access to the pre-trained target model or its original training dataset, the method comprising, at the device:
sending a set of random or semi-random input data to a remote device to randomly probe the pre-trained target model remotely by inputting the set of random or semi-random input data into the pre-trained target model;
receiving from the remote device a set of corresponding output data generated by applying the pre-trained target model to the set of random or semi-random input data;
generating a random probe training dataset comprising the set of random or semi-random input data and corresponding output data generated by randomly probing the pre-trained target model;
training a new model with the random probe training dataset so that the new model generates substantially the same corresponding output data in response to said input data to mimic the pre-trained target model; and
defining data to be omitted from the random probe training dataset to re-train the new model by eliminating a category or class of prediction in the pre-trained target model to unlink input data from corresponding to the eliminated category or class output data such that the unlinked input data cancels out or diminishes to have no average or reduced overall effect on the corresponding output data.