Training machine-learned models with label differential privacy
An example method is provided for conducting differentially private communication of training data for training a machine-learned model. Initial label data can be obtained that corresponds to feature data. A plurality of label bins can be determined to respectively provide representative values for initial label values assigned to the plurality of label bins. Noised label data can be generated, based on a probability distribution over the plurality of label bins, to correspond to the initial label data, the probability distribution characterized by, for a respective noised label corresponding to a respective initial label of the initial label data, a first probability for returning a representative value of a label bin to which the respective initial label is assigned, and a second probability for returning another value. The noised label data can be communicated for training the machine-learned model.
1 . A computer-implemented method for differentially private communication of training data for training a machine-learned model, the method comprising:
obtaining, by a computing system comprising one or more processors, initial label data that corresponds to feature data;
determining, by the computing system, a plurality of label bins respectively providing representative values for initial label values assigned to the plurality of label bins;
generating, by the computing system and based on a probability distribution over the plurality of label bins, noised label data corresponding to the initial label data, the probability distribution characterized by, for a respective noised label corresponding to a respective initial label of the initial label data:
a first probability of returning a representative value of a label bin to which the respective initial label is assigned, and
a second probability of returning another value;
wherein the probability distribution is shaped by a first privacy parameter value; and
communicating, by the computing system, the noised label data for training the machine-learned model.
2 . The computer-implemented method of claim 1 , wherein the other value is a representative value of any other label bin to which the respective initial label is not assigned.
3 . The computer-implemented method of claim 1 , wherein the first probability is related to the first privacy parameter value.
4 . The computer-implemented method of claim 1 , wherein the representative values are determined using an optimization objective that corresponds to expected values of a noise-based loss.
5 . The computer-implemented method of claim 4 , wherein the noise-based loss comprises a regression loss.
6 . The computer-implemented method of claim 4 , wherein the noise-based loss is determined between noised labels and corresponding initial labels.
7 . The computer-implemented method of claim 4 , wherein the optimization objective is computed by weighting respective computed values of the noise-based loss based on probabilities associated with the respective computed values.
8 . The computer-implemented method of claim 7 , wherein the probabilities are obtained from a prior distribution of probabilities associated with the initial label data used to compute the noise-based loss.
9 . The computer-implemented method of claim 8 , comprising:
estimating, by the computing system, the prior distribution of probabilities associated with the initial label data; and
weighting, by the computing system, computed values of the noise-based loss based on the estimated prior distribution of probabilities.
10 . The computer-implemented method of claim 9 , wherein estimating the prior distribution comprises:
determining, by the computing system, a histogram over the initial label data;
injecting, by the computing system, noise into the histogram; and
determining, by the computing system, the estimated prior distribution.
11 . The computer-implemented method of claim 10 , wherein the injected noise is inversely correlated to a second privacy parameter value.
12 . The computer-implemented method of claim 8 , wherein the noised label data is generated by a user computing device, and wherein the prior distribution is generated based on a history of initial label data associated with a user account corresponding to the user computing device.
13 . The computer-implemented method of claim 8 , wherein the noised label data is generated by a user computing device, and wherein the prior distribution is obtained from a remote computing device.
14 . The computer-implemented method of claim 13 , wherein the prior distribution comprises global prior data generated based on global label data associated with a plurality of user devices.
15 . The computer-implemented method of claim 13 , comprising:
submitting, by the computing system and to an application programming interface of the remote computing device, the initial label data.
16 . The computer-implemented method of claim 1 , comprising:
obtaining, by the computing system and from a plurality of user computing devices, the initial label data; and
generating, by the computing system, the noised label data, wherein the noised label data is aggregated across the plurality of user computing devices.
17 . The computer-implemented method of claim 16 , comprising:
transmitting, by the computing system and to a third-party device, the noised label data.
18 . The computer-implemented method of claim 17 , wherein the third-party device has access to the feature data.
19 . A computing system for conducting differentially private communication of training data for training a machine-learned model, the system comprising:
one or more processors; and
one or more non-transitory computer-readable media storing instructions that are executable by the one or more processors to cause the computing system to perform operations, the operations comprising:
obtaining initial label data that corresponds to feature data;
determining a plurality of label bins respectively providing representative values for initial label values assigned to the plurality of label bins;
generating, based on a probability distribution over the plurality of label bins, noised label data corresponding to the initial label data, the probability distribution characterized by, for a respective noised label corresponding to a respective initial label of the initial label data:
a first probability of returning a representative value of a label bin to which the respective initial label is assigned, and
a second probability of returning another value;
wherein the probability distribution is shaped by a first privacy parameter value; and
communicating the noised label data for training the machine-learned model.
20 . One or more non-transitory computer-readable media storing instructions that are executable by one or more processors to cause a computing system to perform operations for conducting differentially private communication of training data for training a machine-learned model, the operations comprising:
obtaining initial label data that corresponds to feature data;
determining a plurality of label bins respectively providing representative values for initial label values assigned to the plurality of label bins;
generating, based on a probability distribution over the plurality of label bins, noised label data corresponding to the initial label data, the probability distribution characterized by, for a respective noised label corresponding to a respective initial label of the initial label data:
a first probability of returning a representative value of a label bin to which the respective initial label is assigned, and
a second probability of returning another value;
wherein the probability distribution is shaped by a first privacy parameter value; and
communicating the noised label data for training the machine-learned model.