Distributed architecture for artificial intelligence model distribution and access control

Patent metadata

Field	Value
Publication number	US-2025272414-A1
Application number	US-202418888657-A
Country	US
Kind code	A1
Filing date	Sep 18, 2024
Priority date	Feb 22, 2024
Publication date	Aug 28, 2025
Grant date	—

Abstract

Aspects of the technology disclosed herein related to a distributed architecture for securely delivering AI models and/or training data sets to client devices for local use. Model creators are able to provide encrypted versions of their models to a centralized server that in turn distributes copies of the encrypted models to a plurality of distribution servers. When a request for a particular model is received from a client device, the best-suited distribution server is identified to deliver the model. The distributed architecture further includes a licensing server that controls access to and decryption of the models. The licensing server controls the distribution of licensing packages for the different models delivered by the distribution server.

First claim

We claim:

1 . A system for distributing and controlling access to artificial intelligence models and training data sets, the system comprising:
a centralized server comprising:
one or more processors; and
a memory storing instructions, that when executed by the one or more processors, cause the centralized server to perform operations comprising:
receive at least one of an encrypted AI model or an encrypted training data set;
distribute the received at least one of the encrypted AI model or encrypted training data set to a plurality of distributed servers, located in different geographic locations, including a first distributed server and a second distributed server;

the first distributed server, of the plurality of distribution servers, comprising:
one or more processors; and
a memory storing instructions, that when executed by the one or more processors, cause the first distributed server to perform operations comprising:
deliver the at least at least one of the encrypted AI model or encrypted training data set to a first subset of client devices having a latency associated with the first distributed server that is less than a latency associated with a second distributed server;

the second distributed server, of the plurality of distribution servers, comprising:
one or more processors; and
a memory storing instructions, that when executed by the one or more processors, cause the second distributed server to perform operations comprising:
deliver the at least at least one of the encrypted AI model or encrypted training data set to a second subset of client devices having a latency associated with the second distributed server that is greater than a latency associated with a second distributed server.

2 . The system of claim 1 , further comprising:
a licensing server comprising:
one or more processors; and
a memory storing instructions, that when executed by the one or more processors, cause the licensing server to perform operations comprising:
receive a license request from a client device in the first subset of client devices, the license request is for a license for local use of the encrypted AI model or encrypted training data set; and
deliver a license package for the model to the client device, wherein the license package includes a decryption key to decrypt the model on the client device and the license for the model.

3 . The system of claim 2 , wherein the operations caused to be performed by the licensing server further comprise:
receive device-level data from the client device; and compare security requirements set forth in a license for the model to the device-level data; wherein delivering the license package is based on the comparison of the security requirements and the device-level data.

4 . The system of claim 1 , wherein the first distributed server is located in a first city and the second distributed server is located in a second city.

5 . The system of claim 4 , wherein a first component of the AI model is encrypted with a first encryption key and a second component of the AI model is encrypted with a second encryption key.

6 . The system of claim 5 , wherein the first component is a network layer topology of the AI model and the second component is model weights of the AI model.

7 . The system of claim 1 , wherein the AI model is one of a deep neural network, a recurrent neural network, a convolution neural network, a small language model, or a large language model.

8 . The system of claim 1 , wherein the first subset of client devices are positioned geographically closer to the first distribution server than the second distribution server.

9 . A system for controlling access to artificial intelligence models, the system comprising:
one or more processors; and a memory storing instructions, that when executed by the one or more processors, cause the system to perform operations comprising:
generating a request to retrieve an AI model for local use on the system;
receiving an identification of a distribution server, storing an encrypted version of the AI model, that is located closest to the system as compared to other distribution servers storing the encrypted version of the AI model;
transmitting the generated request to the identified distribution server;
receiving, from the identified distribution server, the encrypted version of the AI model;
transmitting a license request to a licensing server for a license to use the AI model on the system; and
receiving, from the licensing server, a license package including the license and a decryption key for the model.

10 . The system of claim 9 , wherein the license package includes a first decryption key to decrypt a first component of the model and a second decryption key to decrypt a second component of the model.

11 . The system of claim 9 , wherein the operations further comprise:
generating a request for available AI models; and receiving a list of available AI models, including the AI model.

12 . The system of claim 11 , wherein the request for available AI models includes device-level details of the system, and the list of available AI models is based on the device-level details.

13 . The system of claim 11 , wherein the request for available AI models includes account-level details for a user of the system, and the list of available AI models is based on the account-level details.

14 . The system of claim 9 , wherein the identified distributed server is located geographically closest to the system.

15 . The system of claim 9 , wherein the identified distributed server is located closest to the system based on at least one of network topology, latency, or network transmission cost.

16 . A computer-implemented method for distributing and controlling access to artificial intelligence models, the method comprising:
receiving, by a central server, a plurality of different AI models; distributing, by the central server, the plurality of different AI models a plurality of different distribution servers located in different geographical locations; receiving, by a licensing server, security requirements for the plurality of different AI models; receiving, by the licensing server, decryption keys for the plurality of different AI models; and generating different licensing packages for the plurality of different AI models, wherein each licensing package includes the security requirements for the corresponding AI model and the decryption key for the corresponding AI model.

17 . The method of claim 16 , further comprising:
receiving a request, from a client device, for a particular AI model of the plurality of different models; identifying a particular distribution server, of the plurality of distribution servers, having a lowest latency for communication with the client device; delivering, from the identified distribution server, an encrypted version of the particular AI model to the client device; and delivering, by the licensing server, a particular license package, from the different licensing packages, for the particular AI model.

18 . The method of claim 17 , further comprising precompiling at least a portion of the AI model prior to delivering the AI model to the client device.

19 . The method of claim 16 , wherein the different AI models include models that have at least one of different sizes, different accuracies, or different performance attributes.

20 . The method of claim 16 , wherein the security requirements are different for different

Assignees

• Microsoft Technology Licensing Llc

Inventors

• Rajeev Kumar
• Pathirana Isuru Chamara
• Evans Glenn F
• Stankiewicz Marcin Maciej
• Patel Priya Rakesh

Classifications

• G06F21/60

  Protecting data · CPC title

• G06F21/10 Primary

  Protecting distributed programs or content, e.g. vending or licensing of copyrighted material (protection in video systems or pay television H04N7/16) {; Digital rights management [DRM]} · CPC title

• G06F21/1074

  Definition · CPC title

• G06F21/602 Primary

  Providing cryptographic facilities or services · CPC title