Data protection for remote artificial intelligence models

Patent metadata

Field	Value
Publication number	US-12549347-B2
Application number	US-202117457717-A
Country	US
Kind code	B2
Filing date	Dec 6, 2021
Priority date	Dec 6, 2021
Publication date	Feb 10, 2026
Grant date	Feb 10, 2026

Abstract

A method, computer system, and a computer program product for data protection is provided. The present invention may include, generating an encoder network. The present invention may also include, encoding a training data using the generated encoder network, wherein the training data includes natural language data. The present invention may further include, training a deep learning model using the encoded training data.

First claim

What is claimed is:

1 . A computer-implemented method, comprising:
enhancing data protection of training data in remote training of machine learning (ML) models configured on remote artificial intelligence (AI) platforms, the enhancing comprising:
generating an encoder network according to a set of seeding network parameters;
encoding a training data using the generated encoder network, wherein the training data includes natural language data;
configuring a deep learning (DL) model as an integration of a decoder network and a classifier network, wherein the decoder network comprises a first subset of a set of hidden layers of the DL model and the classifier network comprises a second subset of the set of hidden layers of the DL model; and
executing a training mode of the DL model using the encoded training data such that the training optimizes the decoder network component of the DL model while omitting using the set of seeding network parameters in training the decoder network component of the DL model, the executing outputting a trained DL model.

2 . The method of claim 1 ,
wherein generating the encoder network further comprises generating a user-specific neural network on a local device associated with a user, and wherein training the deep learning model using the encoded training data further comprises transmitting the encoded training data to a remote device storing the deep learning model.

3 . The method of claim 1 , wherein generating the encoder network further comprises:
receiving a unique user key and a current timestamp; and generating a neural network based on the received unique user key and the current timestamp.

4 . The method of claim 3 , wherein generating the neural network based on the received unique user key and the current timestamp further comprises:
initializing the set of seeding network parameters for the generated neural network based on the received unique user key and the current timestamp as seeds, such that the initialized set of seeding network parameters includes a plurality of random numbers.

5 . The method of claim 1 , wherein encoding the training data using the generated encoder network further comprises:
converting the natural language data of the training data to a plurality of embeddings; and mapping the plurality of embeddings from a relatively low-dimensional embedding to a high-dimensional embedding to increase data privacy of the training data.

6 . The method of claim 5 , wherein mapping the plurality of embeddings from the relatively low-dimensional embedding to the high-dimensional embedding further comprises:
determining a set of semantic features captured by the relatively low-dimensional embedding; and maintaining the determined set of semantic features in the high-dimensional embedding.

7 . The method of claim 1 , wherein the training data comprises at least one input and at least one output, and wherein encoding the training data further comprises:
converting the at least one input from the natural language data to an embedding; and maintaining the natural language data of the at least one output.

8 . The method of claim 2 , further comprising:
encoding a prediction request on the local device using the generated user-specific neural network; transmitting the encoded prediction request from the local device to the trained deep learning model stored on the remote device; and receiving, on the local device, a natural language output from the trained deep learning model responsive to the encoded prediction request.

9 . A computer system for data protection, comprising:
one or more processors, one or more computer-readable memories, one or more computer-readable tangible storage media, and program instructions stored on at least one of the one or more computer-readable tangible storage media for execution by at least one of the one or more processors via at least one of the one or more memories, wherein the computer system is capable of performing a method comprising: enhancing data protection of training data in remote training of machine learning (ML) models configured on remote artificial intelligence (AI) platforms, the enhancing comprising:
generating an encoder network according to a set of seeding network parameters;
encoding a training data using the generated encoder network, wherein the training data includes natural language data;
configuring a deep learning (DL) model as an integration of a decoder network and a classifier network, wherein the decoder network comprises a first subset of a set of hidden layers of the DL model and the classifier network comprises a second subset of the set of hidden layers of the DL model; and
executing a training mode of the DL model using the encoded training data such that the training optimizes the decoder network component of the DL model while omitting using the set of seeding network parameters in training the decoder network component of the DL model, the executing outputting a trained DL model.

10 . The computer system of claim 9 ,
wherein generating the encoder network further comprises generating a user-specific neural network on a local device associated with a user, and wherein training the deep learning model using the encoded training data further comprises transmitting the encoded training data to a remote device storing the deep learning model.

11 . The computer system of claim 9 , wherein generating the encoder network further comprises:
receiving a unique user key and a current timestamp; and generating a neural network based on the received unique user key and the current timestamp.

12 . The computer system of claim 11 , wherein generating the neural network based on the received unique user key and the current timestamp further comprises:
initializing the set of seeding network parameters for the generated neural network based on the received unique user key and the current timestamp as seeds, such that the initialized set of seeding network parameters includes a plurality of random numbers.

13 . The computer system of claim 9 , wherein encoding the training data using the generated encoder network further comprises:
converting the natural language data of the training data to a plurality of embeddings; and mapping the plurality of embeddings from a relatively low-dimensional embedding to a high-dimensional embedding to increase data privacy of the training data.

14 . The computer system of claim 13 , wherein mapping the plurality of embeddings from the relatively low-dimensional embedding to the high-dimensional embedding further comprises:
determining a set of semantic features captured by the relatively low-dimensional embedding; and maintaining the determined set of semantic features in the high-dimensional embedding.

15 . The computer system of claim 9 , wherein the training data comprises at least one input and at least one output, and wherein encoding the training data further comprises:
converting the at least one input from the natural language data to an embedding; and maintaining the natural language data of the at least one output.

16 . The computer system of claim 10 , further comprising:
encoding a prediction request on the local device using the generated user-specific neural network; transmitting the encoded prediction request from the local device to the trained deep learning model stored on the remote device; and receiving, on the local device, a natural language output from the trained deep learning model responsive to the encoded prediction request.

Assignees

• IBM

Inventors

• Gao Li Juan
• Yuan Zhong Fang
• Chen Ming Jin
• Liu Tong

Classifications

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06V30/19147

  Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06F40/40

  Processing or translation of natural language (natural language analysis G06F40/20; semantic analysis G06F40/30) · CPC title

• H04L9/0819

  Key transport or distribution, i.e. key establishment techniques where one party creates or otherwise obtains a secret value, and securely transfers it to the other(s) (network architectures or network communication protocols for key distribution in a packet data network H04L63/062) · CPC title