Deep learning-based detection and data loss prevention of image-borne sensitive documents

Patent metadata

Field	Value
Publication number	US-11537745-B2
Application number	US-202017116862-A
Country	US
Kind code	B2
Filing date	Dec 9, 2020
Priority date	Jun 3, 2020
Publication date	Dec 27, 2022
Grant date	Dec 27, 2022

Abstract

The technology disclosed relates to distributing a trained master deep learning (DL) stack with stored parameters to a plurality of organizations, to detect organization sensitive data in images, referred to as image-borne organization sensitive documents, and protecting against loss of the image-borne organization sensitive documents. Disclosed is providing organizations with a DL stack update trainer, under the organizations' control, configured to allow the organizations to perform update training to generate updated DL stacks, without the organizations forwarding images of organization-sensitive training examples, and to save non-invertible features derived from the images, ground truth labels for the images, and parameters of the updated DL stacks. In particular, the technology disclosed relates to receiving, from a plurality of the DL stack update trainers, organization-specific examples including the non-invertible features of the organization-sensitive training examples and the ground truth labels, and using the received organization-specific examples to update the trained master DL stack.

First claim

What is claimed is:

1. A computer-implemented method of customizing a deep learning (abbreviated DL) stack to detect organization sensitive data in images, referred to as image-borne organization sensitive documents, and protecting against loss of the image-borne organization sensitive documents, including:
distributing a trained master DL stack with stored parameters to a plurality of organizations;
providing at least some of the organizations with a DL stack update trainer, under the organizations' control, configured to allow the organizations to perform update training to generate updated DL stacks, without the organizations forwarding images of organization-sensitive training examples, and to save non invertible features derived from the images, ground truth labels for the images, and parameters of the updated DL stacks;
receiving, from a plurality of the DL stack update trainers, organization-specific examples including the non-invertible features of the organization-sensitive training examples and the ground truth labels; and
using the received organization-specific examples to update the trained master DL stack.

2. The computer-implemented method of claim 1 , further including:
directing the organization-sensitive training examples to training of a second set of layers of the trained master DL stack, overlaying a general image processing first set of layers;
storing updated parameters of the second set of layers for inference from production images; and
distributing the updated parameters of the second set of layers to the plurality of organizations.

3. The computer-implemented method of claim 1 , further including performing from-scratch re-training, with the organization-sensitive training examples and non-invertible features from other examples used to train the trained master DL stack, to re-train the trained master DL stack.

4. A tangible non-transitory computer readable storage media, including program instructions loaded into memory that, when executed on processors, cause the processors to implement a computer-implemented method of customizing a deep learning (abbreviated DL) stack to detect organization sensitive data in images, referred to as image-borne organization sensitive documents, and protecting against loss of the image-borne organization sensitive documents, the computer-implemented method including:
distributing a trained master DL stack with stored parameters to a plurality of organizations;
providing at least some of the organizations with a DL stack update trainer, under the organizations' control, configured to allow the organizations to perform update training to generated updated DL stacks, without the organizations forwarding images of organization-sensitive training examples, and to save non invertible features derived from the images, ground truth labels for the images, and parameters of the updated DL stacks;
receiving, from a plurality of the DL stack update trainers, organization-specific examples including the non-invertible features of the organization-sensitive training examples and the ground truth labels; and
using the received organization-specific examples to update the trained master DL stack.

5. The tangible non-transitory computer readable storage media of claim 4 , further implementing:
directing the organization-sensitive training examples to training of a second set of layers of the trained master DL stack, overlaying a general image processing first set of layers;
storing updated parameters of the second set of layers for inference from production images; and
distributing the updated parameters of the second set of layers to the plurality of organizations.

6. The tangible non-transitory computer readable storage media of claim 4 , further implementing from-scratch re-training, with the organization-sensitive training examples and non-invertible features from other examples used to train the trained master DL stack, to re-train the trained master DL stack.

7. A system for customizing a deep learning (abbreviated DL) stack including a processor, memory coupled to the processor, and computer instructions from the non-transitory computer readable storage media of claim 4 loaded into the memory.

8. A system for customizing a deep learning (abbreviated DL) stack including a processor, memory coupled to the processor, and computer instructions from the non-transitory computer readable storage media of claim 5 loaded into the memory.

9. A system for customizing a deep learning (abbreviated DL) stack including a processor, memory coupled to the processor, and computer instructions from the non-transitory computer readable storage media of claim 6 loaded into the memory.

10. A computer-implemented method of customizing a deep learning (abbreviated DL) stack to detect organization sensitive data in images, referred to as image-borne organization sensitive documents, and protecting against loss of the image-borne organization sensitive documents, including:
distributing a trained master DL stack with stored parameters to a plurality of organizations;
providing at least some of the organizations with a DL stack update trainer, under the organizations' control, configured to allow the organizations to perform update training to generate updated DL stacks, without the organizations forwarding images of organization-sensitive training examples, and to save non invertible features derived from the images, labels for the images, and parameters of the updated DL stacks;
receiving, from a plurality of the DL stack update trainers, organization-specific parameters of updated DL stacks; and
using the received organization-specific parameters of updated DL stacks to update the trained master DL stack.

11. The computer-implemented method of claim 10 , further including:
the DL stack update trainers configured to forward updated coefficients from a second set of layers, overlaying a general image processing first set of layers;
receiving from a plurality of the DL stack update trainers, respective updated coefficients from respective second sets of layers;
combining the updated coefficients from respective second sets of layers to further train the second set of layers of the trained master DL stack;
storing updated parameters of the second set of layers of the trained master DL stack for inference from production images; and
distributing the updated parameters of the second set of layers to the plurality of organizations.

12. The computer-implemented method of claim 10 , further including:
storing updated parameters of a second set of layers for inference from production images; and
distributing the updated parameters of the second set of layers to the plurality of organizations.

13. A tangible non-transitory computer readable storage media, including program instructions loaded into memory that, when executed on processors, cause the processors to implement a computer-implemented method of customizing a deep learning (abbreviated DL) stack to detect organization sensitive data in images, referred to as image-borne organization sensitive documents, and protecting against loss of the image-borne organization sensitive documents, the computer-implemented method including:
distributing a trained master DL stack with stored parameters to a plurality of organizations;
providing at least some of the organizations with a DL stack update trainer, under the organizations' control, configured to allow the organizations to perform update training to generate updated DL stacks, without the organizations forwarding images of organization-sensitive training examples, and to save non invertible features derived fr

Assignees

• Netskope Inc

Inventors

• Yang Siying
• Wang Xiaolin
• Narayanaswamy Krishna
• Zhang Yi

Classifications

• G06N20/00

  Machine learning · CPC title

• G06F21/604

  Tools and structures for managing or administering access control systems · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06F21/6245 Primary

  Protecting personal data, e.g. for financial or medical purposes · CPC title

• G06N3/045

  Combinations of networks · CPC title