Back to results
US 12,719,678 B2Granted
Training neural networks with non-polynomial elements for homomorphic encryption computations using sub-networks and multi-loss
AssigneeInternational Business Machines Corporation
InventorsItamar Zimerman, Nir Drucker
Life of this patent
Estimating the term…
Abstract
An example system includes a processor to receive a non-homomorphic encryption (HE)-friendly analytics model including a non-polynomial element. The processor is to train a substitution model in which the non-polynomial element of the non-homomorphic encryption (HE)-friendly analytics model is replaced with a sub-network including a polynomial replacement element.
Claims19
A computer system comprising:
- a processor set;
- and
- one or more computer-readable storage media;
- wherein:
- the processor set is structured, located, connected, and/or programmed to run program instructions stored on the one or more computer-readable storage media;
- and
- the program instructions which, when executed by the processor set, cause the processor set to perform operations comprising:
- receiving a non-homomorphic encryption (HE)-friendly analytics model comprising a plurality of non-polynomial elements contained in corresponding non-polynomial models, the corresponding non-polynomial models in respectively corresponding layers of the non-HE-friendly analytics model;
- establishing a sub-network for a first layer of a substitution analytics model, the sub-network including a pair of models comprising a polynomial substitution model, in which a non-polynomial element of a corresponding non-polynomial model is replaced with a polynomial replacement element, and the corresponding non-polynomial model;
- training the polynomial substitution model of the sub-network by:
- calculating an imitation loss between outputs of the pair of models, with the non-polynomial model operating as a teacher for the polynomial substitution model;
- and
- outputting a trained substitution analytics model including the trained polynomial substitution model of the sub-network.
10 more elements
A computer-implemented method, comprising:
- receiving a non-homomorphic encryption (HE)-friendly analytics model comprising a plurality of non-polynomial elements contained in corresponding non-polynomial models, the corresponding non-polynomial models in respectively corresponding layers of the non-HE-friendly analytics model;
- establishing a sub-network for a first layer of a substitution analytics model, the sub-network including a pair of models comprising a polynomial substitution model, in which a non-polynomial element of a corresponding non-polynomial model is replaced with a polynomial replacement element, and the corresponding non-polynomial model;
- training the polynomial substitution model of the sub-network by:
- calculating an imitation loss between outputs of the pair of models, with the non-polynomial model operating as a teacher for the polynomial substitution model;
- and
- outputting a trained substitution analytics model including the trained polynomial substitution model of the sub-network.
3 more elements
A computer program product comprising one or more computer-readable storage media having a set of instructions stored therein which, when executed by a processor, causes the processor to perform a method comprising:
- receiving a non-homomorphic encryption (HE)-friendly analytics model comprising a plurality of non-polynomial elements contained in corresponding non-polynomial models, the corresponding non-polynomial models in respectively corresponding layers of the non-HE-friendly analytics model;
- establishing a sub-network for a first layer of a substitution analytics model, the sub-network including a pair of models comprising a polynomial substitution model, in which a non-polynomial element of a corresponding non-polynomial model is replaced with a polynomial replacement element, and the corresponding non-polynomial model;
- training the polynomial substitution model of the sub-network by:
- calculating an imitation loss between outputs of the pair of models, with the non-polynomial model operating as a teacher for the polynomial substitution model;
- and
- outputting a trained substitution analytics model including the trained polynomial substitution model of the sub-network.
3 more elements
Dependencies are inferred from the claim text.