Privacy-enhanced training and deployment of machine learning models using client-side and server-side data
Computer-implemented systems and methods for training a decentralized model for making a personalized recommendation. In one aspect, the method comprising: obtaining, using user activity data, client-side training data that includes features and training labels; and training, by the client device, a decentralized model in training rounds, wherein training, in each training round comprises: receiving, first data including a current server-side embedding generated by the server-side machine learning model, wherein the first data received from the server does not include any server-side data used in generating the current server-side embedding; generating, using the client-side machine learning model, a client-side embedding based on the client-side training data; updating, using the client-side embedding and the current server-side embedding and based on the training labels, the client-side machine learning model; generating, an updated client-side embedding; and transmitting second data including the updated client-side embedding for subsequent updating of the server-side machine learning model.
1 . A computer-implemented method, comprising:
obtaining, by a client device and using user activity data stored at the client device, client-side training data that includes a plurality of features and a corresponding plurality of training labels; and
training, by the client device, a decentralized model in a plurality of training rounds, wherein the decentralized model (1) is trained to output data indicating whether one or more digital components should be provided to the client device based on input user activity data and digital component attribute data, (2) includes a client-side machine learning model and a server-side machine learning model, and (3) is trained while ensuring joint differential privacy, wherein:
the server-side machine learning model comprises at least a first sub-model corresponding to a content provider and a second sub-model corresponding to a content publisher;
each of the first sub-model and the second sub-model are configured to generate a respective server-side embedding; and
training, by the client device, in each training round of the plurality of training rounds comprises:
receiving, from a server, first data including a current server-side embedding generated by the server-side machine learning model, wherein the first data received from the server does not include any server-side features used in generating the current server-side embedding;
updating, using current server-side embedding and based on the plurality of training labels, the client-side machine learning model;
generating, using the updated client-side machine learning model, an updated client-side embedding; and
transmitting, by the client device and to the server, second data including the updated client-side embedding for subsequent updating of the server-side machine learning model, wherein updating the client-side machine learning model comprises minimizing a client-side loss function based on a combination of the current server-side embedding and a client-side embedding while parameters of the server-side machine learning model are held fixed.
2 . The computer-implemented method of claim 1 , wherein:
the second data transmitted to the server does not include any of the plurality of features of the client-side training data used in generating the updated client-side embedding.
3 . The computer-implemented method of claim 1 ,
wherein training, by the client device, in each training round of the plurality of training rounds comprises:
combining the client-side embedding and the current server-side embedding to obtain a combined embedding, and
wherein updating, using the client-side embedding and the current server-side embedding and based on the plurality of training labels, the client-side machine learning model comprises
updating, using the combined embedding and based on the plurality of training labels, the client-side machine learning model.
4 . The computer-implemented method of claim 3 , wherein updating the client-side machine learning model comprises:
minimizing a loss function based on the combined embedding while keeping parameters of the server-side machine learning model unaltered.
5 . The computer-implemented method of claim 4 , wherein minimizing the loss function is based on stochastic gradient descent or alternating least squares.
6 . The computer-implemented method of claim 1 , wherein the client-side machine learning model is shared among a plurality of client devices including the client device and a second client device based on joint differential privacy.
7 . The computer-implemented method of claim 1 , wherein the client-side machine learning model is independent from a second client-side machine learning model for the second client device.
8 . The computer-implemented method of claim 7 , wherein the client-side training data and the input user activity data is not shared with or accessible by the second client device.
9 . The computer implemented method of claim 7 , wherein combining the client-side embedding and the current server-side embedding to obtain a combined embedding is by using a dot product or a tensor product.
10 . A computer-implemented method, comprising:
obtaining, by a server and using server-side data obtained from a server-side user, server-side training data that includes a plurality of features and a corresponding plurality of training labels; and
training, by the server, at least part of a decentralized model in a plurality of training rounds, wherein the decentralized model (1) is trained to output data indicating whether one or more digital components should be provided to the client device based on input user activity data and digital component attribute data, and (2) includes a client-side machine learning model and a server-side machine learning model,
wherein:
the server-side machine learning model comprises at least a first sub-model corresponding to a content provider and a second sub-model corresponding to a content publisher;
each of the first sub-model and the second sub-model are configured to generate a respective server-side embedding; and
training, by the server, in each training round of the plurality of training rounds comprises:
receiving, from a client device, second data including a current client-side embedding generated by the client-side machine learning model, wherein the second data received from the client device does not include any client-side features used in generating the current client-side embedding;
updating, using server-side embedding and the current client-side embedding and based on the plurality of training labels, the server-side machine learning model;
generating, using the updated server-side machine learning model, an updated server-side embedding; and
transmitting, by the server and to the client device, first data including the updated server-side embedding for subsequent updating of the client-side machine learning model, wherein updating the server-side machine learning model comprises minimizing a server-side loss function based on a combination of the current client-side embedding and a server-side embedding while parameters of the client-side machine learning model are held fixed.
11 . The computer-implemented method of claim 10 , wherein
the first data transmitted to the client device does not include any of the plurality of features of the server-side training data used in generating the updated server-side embedding.
12 . The computer-implemented method of claim 10 ,
wherein the training, by the server, in each training round of the plurality of training rounds comprises:
combining the current client-side embedding and the server-side embedding to obtain a combined embedding, and
wherein the updating, using the current client-side embedding and the server-side embedding and based on the plurality of training labels, the server-side machine learning model comprises:
updating, using the combined embedding and based on the plurality of training labels, the server-side machine learning model.
13 . The computer-implemented method of claim 12 , wherein updating the server-side machine learning model comprises:
minimizing a server loss function based on the combined embedding while keeping parameters of the client-side machine learning model unaltered.
14 . The computer-implemented method of claim 13 , wherein minimizing the server loss function is based on stochastic gradient descent or alternating least squares.
15 . The computer-implemented method of claim 10 , wherein the client-side machine learning model is shared among a plurality of client devices including the client device and a second client device based on joint differential privacy.
16 . The computer-implemented method of claim 15 , wherein the client-side machine learning model is independent from a second client-side machine learning model for the second client device.
17 . The computer-implemented method of claim 16 , wherein client-side training data and the input user activity data is not shared with or accessible by the second client device.
18 . The computer-implemented method of claim 10 , wherein the server-side user is a content publisher or a content provider, and wherein the server-side training data includes training digital component attribute data.
19 . The computer-implemented method of claim 12 , wherein combining the current client-side embedding and the server-side embedding to obtain a combined embedding is by using a dot product or a tensor product.
20 . A computer-implemented method performed by a server, comprising:
receiving, from a client device, a request for a digital component to be provided for display within a content page displayed on the client device;
receiving, from the client device, first data including a client-side embedding generated by a trained client-side machine learning model, wherein;
the server-side machine learning model comprises at least a first sub-model corresponding to a content provider and a second sub-model corresponding to a content publisher;
each of the first sub-model and the second sub-model are configured to generate a respective server-side embedding; and
the trained client-side machine learning model and a server-side machine learning model jointly define a decentralized model trained to output data indicating whether one or more digital components should be provided to a client device based on input user activity data and digital component attribute data;
generating, using the server-side machine learning model and based on digital component attribute data relating to a plurality of digital components, a server-side embedding, wherein the digital component attribute data is stored at the server;
combining the server-side embedding and the client-side embedding to obtain a combined embedding;
generating, using the combined embedding and for each digital component in the plurality of digital components, a score indicating whether the digital component should be provided to the client device;
selecting, from among the plurality of digital components and based on the scores for the plurality of digital components, a particular digital component; and
transmitting, to the client device, the particular digital component for provision within the content page displayed on the client device.
21 . The computer-implemented method of claim 20 , wherein:
selecting, from among the plurality of digital components and based on the scores for the plurality of digital components, the particular digital component includes selecting the particular digital component in response to determining that the score corresponding to the particular digital component is greater than scores corresponding to other digital components in the plurality of digital components.
22 . The computer-implemented method of claim 20 , wherein the combining the server-side embedding and the client-side embedding to obtain the combined embedding comprises:
combining the server-side embedding and the client-side embedding using a dot product or a tensor product.
23 . The computer-implemented method of claim 20 , wherein the first data does not include any user activity data used in generating the client-side embedding.
24 . The computer-implemented method of claim 20 , wherein the server includes a plurality of servers, and wherein each server in the plurality of servers includes a respective server-side machine learning model for generating a respective server-side embedding.
25 . The computer-implemented method of claim 20 , wherein combining the server-side embedding and the client-side embedding to obtain the combined embedding includes:
combining the client-side embedding with a plurality of client side embeddings received from a plurality of other client devices to obtain an aggregated client-side embedding; and
combining the aggregated client-side embedding and the server-side embedding to obtain the combined embedding.