Amazon patents privacy-protecting method for training image recognition models
This patent describes a way to train machine learning models that recognize images while protecting the privacy of the data used. It limits how much any single image can change the model's learning process, adjusting these limits dynamically based on previous steps to balance privacy and accuracy.
In the filing’s own words
Methods and systems for preserving privacy and maintaining performance of machine learning models by limiting an amount of noise and data that are added by an individual member of a training dataset to a machine learning model being trained are disclosed. In some embodiments, the limiting of the noise and influence of an individual members is adaptively adjusted when performing the machine learning model training according to privacy needs of a set of individual members of the training dataset in a differentially private training system. Such techniques may be used in open-classification machine learning model training.
What exactly they claim (claim 1)
1. A system, comprising, a memory storing program instructions; and one or more processors, wherein the program instructions, when executed on or across the one or more processors, cause the one or more processors to: train an open-classification machine learning model using a set of training data comprising a set of publicly available images and a set of non-publicly available images, wherein said training of the open-classification machine learning model is performed via a plurality of partial training iterations, wherein an individual iteration of partial training uses one or more individual members of the set of publicly available images and one or more individual members of the set of non-publicly available images; and adjust an influence value used in the plurality of partial training iterations, wherein said adjustment of the influence value is limited to a threshold level of adjustment based on a prior influence value of a prior iteration of partial training of said training of the open-classification machine learning model, and wherein the influence value: is based on one or more gradients of the one or more of the individual members of the set of publicly available images the individual iteration of partial training is using; and limits an effect of the one or more of the individual members of the set of non-publicly available images on a weight for a loss amount of the individual iteration of partial training.
Facts
- Company
- Amazon (Amazon Technologies, Inc.)
- Official title
- Differentially private open-classification machine learning model training with adaptive clipping
- Patent number
- US12755517B1
- Application number
- 18536040
- Filed
- Dec 11, 2023
- Granted
- Oct 6, 2026
- Inventors
- 7
- Claims
- 20
Classified under
G06V 10/774 Physics & computingG06V 10/776 Physics & computingG06V 40/168 Physics & computingG06V 40/172 Physics & computingMore from Amazon in AI & software
- Amazon patents system that cites sources for AI answers Sep 29, 2026
- Amazon patents tech to swap words in movies using actor's voice Sep 8, 2026
- Amazon patents ML tool to add graphics to live sports from moving cameras Sep 15, 2026
- Amazon patents shopping assistant that remembers your search filters Oct 6, 2026
- Amazon patents system where AI agents split tasks and talk to each other Oct 20, 2026
Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.