- Number
- 12387004
- Published
- 2025-08-12
- Filed
- 2022-03-24
- Assignee
- Disney Enterprises, Inc.
- Inventors
- Farre Guiu; Miquel Angel et al.
- CPC
- G06N3/096; G06F21/6254; G06N3/08; G06N3/045
- Verdict
- Set aside content anonymization, privacy/business
- Source
- Google Patents · FreePatentsOnline
Abstract
A system includes a computing platform having processing hardware, and a system memory storing software code and a machine learning (ML) model. The processing hardware is configured to execute the software code to receive from a client, a request for a dataset, the request identifying a content type of the dataset, obtain the dataset having the content type, and select, based on the content type, an anonymization technique for the dataset, the anonymization technique selected so as to render at least one feature included in the dataset recognizable but unidentifiable. The processing hardware is further configured to execute the software code to anonymize, using the ML model and the selected anonymization technique, the at least one feature included in the dataset, and to output to the client, in response to the request, an anonymized dataset including the at least one anonymized feature.
Background
BACKGROUND (1) There are many undertakings that can be advanced more efficiently when performed collaboratively. However, due to any of potentially myriad privacy or confidentiality concerns, it may be undesirable to share data including proprietary or otherwise sensitive information in a collaborative or other semi-public or public environment. (2) By way of example, collaborative machine learning development endeavors, such as “hackathons” for instance, can advantageously accelerate the process of identifying and optimizing machine learning models for use in a variety of applications, such as activity recognition, location recognition, facial recognition, and object recognition. However, due to the proprietary nature or sensitivity of certain types of content, it may be undesirable to make such content generally available for use in model training. (3) One conventional approach to satisfying the competing interests of content availability and content security for machine learning development is to utilize a remote execution platform as a privacy shield between model developers and content owners. According to this approach, the remote execution platform can mediate training of a machine learning model using proprietary content, while sequestering that content from the model developer. However, one disadvantage of this approach is that, because the model developer is prevented from accessing the content used for training, it is difficult or impossible to accurately assess th
Claims
1. A system comprising: a processing hardware; and a system memory storing a software code, a first trained neural network (NN) and a second trained NN; the processing hardware configured to execute the software code to: receive from a client, a request for a dataset, the request identifying a content type of the dataset, the dataset including an image; obtain the dataset having the content type; select, based on the content type, an anonymization technique for the dataset, the anonymization technique selected so as to render at least one feature included in the dataset recognizable but unidentifiable; anonymize, using the first trained NN and the selected anonymization technique, the at least one feature included in the image, such that a generic nature of the at least one feature included in the image is maintained by anonymizing, but an identifiable nature of the at least one feature included in the image is removed by anonymizing, wherein the at least one feature included in the image and anonymized is indicative of at least one of a location where the image is captured or an activity being performed, wherein the activity includes one of dancing, fencing, first bumping, hand-to-hand fighting, hand clapping, handshaking, hand-fiving, holding hands, hugging, performing magic, or pushing; evaluate, using the second trained NN, an anonymity of the at least one anonymized feature; re-anonymize, using the first trained NN, the at least one anonymized feature when evaluating the anonymity of the at least one anonymized feature fails to confirm that the at least one anonymized feature is unidentifiable, until when evaluating the anonymity of the at least one anonymized feature indicates that a confidence value of the at least one anonymized feature being unidentifiable satisfies a predetermined threshold; and output to the client, in response to the request, an anonymized dataset including the image having the at least one anonymized feature. ||
7. A method for use by a system including a processing hardware, and a system memory storing a software code, a first trained neural network (NN) and a second trained NN, the method comprising: receiving from a client, by the software code executed by the processing hardware, a request for a dataset, the request identifying a content type of the dataset, the dataset including an image; determining, by the software code executed by the processing hardware, quota for content having the content type does not exceed an allowable limit; obtaining, by the software code executed by the processing hardware, the dataset having the content type and including the content; selecting, by the software code executed by the processing hardware, based on the content type, an anonymization technique for the dataset, the anonymization technique selected so as to render at least one feature included in the dataset recognizable but unidentifiable; anonymizing, by the software code executed by the processing hardware and using the first trained NN and the selected anonymization technique, the at least one feature included in the image, such that a generic nature of the at least one feature included in the image is maintained by anonymizing, but an identifiable nature of the at least one feature included in the image is removed by anonymizing, wherein the at least one feature included in the image and anonymized is indicative of at least one of a location where the image is captured or an activity being performed, wherein the activity includes one of dancing, fencing, first bumping, hand-to-hand fighting, hand clapping, handshaking, hand-fiving, holding hands, hugging, performing magic, or pushing; evaluate, by the software code executed by the processing hardware and using the second trained NN, an anonymity of the at least one anonymized feature; re-anonymize, by the software code executed by the processing hardware and using the first trained NN, the at least one anonymized feature when evaluating the anonymity of the at least one anonymized feature fails to confirm that the at least one anonymized feature is unidentifiable, until when evaluating the anonymity of the at least one anonymized feature indicates that a confidence value of the at least one anonymized feature being unidentifiable satisfies a predetermined threshold; and outputting to the client, by the software code executed by the processing hardware in response to the request, an anonymized dataset including the image having the at least one anonymized feature.