Method to improve training of classifiers when using data with personal identifiable information
An approach for managing privacy-relevant data. Disclosed embodiments significantly improve the computational efficiency of training machine-learning models while still protecting privacy-relevant data in the training data.
1 . A method for managing privacy-relevant data, comprising:
for data elements in a set of training data elements containing privacy-relevant data, separating a privacy-relevant portion of each data element from a remaining privacy-nonrelevant portion of each data element, wherein the separating is performed in one or more privacy-secure computing environments configured to securely protect the privacy-relevant portion of each data element;
generating, in one or more computationally-efficient computing environments that are more computationally efficient than the one or more privacy-secure computing environments, a first set of training data, wherein the generating of the first set of training data includes labeling one or more items in the privacy-nonrelevant portion of each data element;
training, in the one or more computationally-efficient computing environments, a machine-learning model on the first set of training data;
generating, in the one or more privacy-secure computing environments, a second set of training data, wherein generating of the second set of training data includes labeling one or more items in the privacy-relevant portion of each data element to obtain one or more labeled items and adding the one or more labeled items to the corresponding data element to form the second set of training data used for retraining the machine-learning model; and
retraining, in the one or more privacy-secure computing environments, the machine-learning model, wherein the retraining includes applying a transfer-learning technique with the second set of training data.
2 . The method of claim 1 , wherein the one or more privacy-secure computing environments includes a trusted execution environment (“TEE”), a Privacy Enhancing Technology (“PET”), or a Secure Multi-party Computation (“SMPC”).
3 . The method of claim 1 , wherein the privacy-relevant portion of at least one data element includes a plurality of items of privacy-relevant information.
4 . The method of claim 1 , wherein at least one data element comprises at least one of image data, audio data, sensor data, and Global Positioning System (“GPS”) data.
5 . The method of claim 4 , wherein each data element comprises at least one of image data, audio data, sensor data, and Global Positioning System (“GPS”) data.
6 . The method of claim 1 , wherein the machine-learning model comprises a neural network.
7 . The method of claim 1 , wherein the neural network comprises a convolutional encoder-decoder.
8 . The method of claim 1 , further comprising:
prior to labeling items in the privacy-nonrelevant portion of each data element, combine the privacy-nonrelevant portion of each data element with replacement privacy-nonrelevant data to replace the separated privacy-relevant portion with the replacement privacy-nonrelevant data.
9 . The method of claim 8 , wherein the replacement privacy-nonrelevant data comprises artificial data.
10 . The method of claim 9 , wherein the artificial data is a generated image of a face.
11 . The method of claim 9 , wherein the artificial data is output from a Generative Adversarial Network (“GAN”) or a transformer technique.
12 . The method of claim 11 , wherein the output is a generated image of a face.
13 . A non-transitory memory comprising processor-executable instructions that when executed by one or more processors causes a system to perform operations including:
for data elements in a set of training data elements containing privacy-relevant data, separating a privacy-relevant portion of each data element from a remaining privacy-nonrelevant portion of each data element, wherein the separating is performed in one or more privacy-secure computing environments configured to securely protect the privacy-relevant portion of each data element;
generating, in one or more computationally-efficient computing environments that are more computationally efficient than the one or more privacy-secure computing environments, a first set of training data, wherein the generating of the first set of training data includes labeling one or more items in the privacy-nonrelevant portion of each data element;
training, in the one or more computationally-efficient computing environments, a machine-learning model on the first set of training data;
generating, in the one or more privacy-secure computing environments, a second set of training data, wherein generating of the second set of training data includes labeling one or more items in the privacy-relevant portion of each data element to obtain one or more labeled items and adding the one or more labeled items to the corresponding data element to form the second set of training data used for retraining the machine-learning model; and
retraining, in the one or more privacy-secure computing environments, the machine-learning model, wherein the retraining includes applying a transfer-learning technique with the second set of training data.
14 . The non-transitory memory of claim 13 , wherein:
wherein each data element comprises at least one of image data, audio data, sensor data, and Global Positioning System (“GPS”) data and wherein the machine-learning model comprises a neural network.
15 . The non-transitory memory of claim 14 , wherein the neural network comprises a convolutional encoder-decoder.
16 . The non-transitory memory of claim 15 , wherein the operations further include:
prior to labeling items in the privacy-nonrelevant portion of each data element, combining the privacy-nonrelevant portion of each data element with replacement privacy-nonrelevant data to replace the separated privacy-relevant portion with the replacement privacy-nonrelevant data.
17 . A system, comprising:
one or more processors; and
non-transitory memory communicatively connected to the one or more processors, the non-transitory memory comprising processor-executable instructions that when executed by the one or more processors causes the system to perform operations including:
for data elements in a set of training data elements containing privacy-relevant data, separating a privacy-relevant portion of each data element from a remaining privacy-nonrelevant portion of each data element, wherein the separating is performed in one or more privacy-secure computing environments configured to securely protect the privacy-relevant portion of each data element;
generating, in one or more computationally-efficient computing environments that are more computationally efficient than the one or more privacy-secure computing environments, a first set of training data, wherein the generating of the first set of training data includes labeling one or more items in the privacy-nonrelevant portion of each data element;
training, in the one or more computationally-efficient computing environments, a machine-learning model on the first set of training data;
generating, in the one or more privacy-secure computing environments, a second set of training data, wherein generating of the second set of training data includes labeling one or more items in the privacy-relevant portion of each data element to obtain one or more labeled items and adding the one or more labeled items to the corresponding data element to form the second set of training data used for retraining the machine-learning model; and
retraining, in the one or more privacy-secure computing environments, the machine-learning model, wherein the retraining includes applying a transfer-learning technique with the second set of training data.
18 . The system of claim 17 , wherein:
wherein the one or more privacy-secure computing environments includes a trusted execution environment (“TEE”), a Privacy Enhancing Technology (“PET”), or a Secure Multi-Party Computation (“SMPC”).
19 . The system of claim 18 , wherein the neural network comprises a convolutional encoder-decoder.
20 . The system of claim 19 , wherein the operations further include:
prior to labeling items in the privacy-nonrelevant portion of each data element, combining the privacy-nonrelevant portion of each data element with replacement privacy-nonrelevant data to replace the separated privacy-relevant portion with the replacement privacy-nonrelevant data.