Edge device machine learning
Techniques are disclosed in which a computing device repeatedly trains, using a stream of user data received at the computing device, a baseline model to generate a device-trained model, wherein the baseline model is trained at the computing device without providing user data included in the stream to a server computer system. In some embodiments, the computing device inputs, to the device-trained model, a set of characteristics associated with a user request received from a user of the computing device, wherein the device-trained model outputs a score for the user request. In some embodiments, the computing device transmits, to the server computer system, the score for the user request, wherein the transmitting includes requesting a decision for the user request. In some embodiments, the computing device performs an action associated with the user request in response to receiving a decision for the user request from the server computer system.
1. A method, comprising:
receiving, by a computing device from a server computer system, a baseline model that is generated by the server computer system based on multiple different models;
repeatedly retraining, by the computing device using a stream of user data received at the computing device, the baseline model to generate a device-trained model;
transmitting, by the computing device to the server computer system, the device-trained model and a set of obfuscated user data generated at the computing device from the stream of user data, wherein the transmitting is performed by the computing device utilizing a first type of wireless communication that provides a higher network performance than a secondary type of wireless communication available to the computing device;
in response to the transmitting, receiving, by the computing device from the server computer system, an updated version of the device-trained model;
inputting, by the computing device to the device-trained model, a set of characteristics associated with a user request received from a user of the computing device, wherein the device-trained model outputs a score for the user request;
transmitting, by the computing device to the server computer system, the score for the user request, wherein the transmitting includes requesting a decision for the user request; and
in response to receiving a decision for the user request generated by the server computer system based on the set of obfuscated user data, the computing device performing an action associated with the user request.
2. The method of claim 1 , wherein the stream of user data includes a plurality of characteristics associated with the computing device and the set of characteristics associated with the user request, and wherein the baseline model and the device-trained model are machine learning models.
3. The method of claim 1 , wherein the baseline model is trained by the server computer system without the stream of user data securely stored at the computing device.
4. The method of claim 1 , wherein the repeatedly retraining includes:
generating an aggregated depiction of data included in the stream of user data; and
comparing different portions of the aggregated depiction with one another, wherein the different portions are selected from different intervals of time within the aggregated depiction; and
adjusting the stream of user data based on one or more portions having data that differ a threshold amount.
5. The method of claim 1 , wherein the repeatedly retraining includes:
performing one or more feature engineering techniques on a plurality of characteristics included in the stream of user data, wherein the one or more feature engineering techniques are performed according to one or more conditions of the following conditions: on-the-fly processing, lookup-based processing, and cache-based processing.
6. The method of claim 1 , further comprising:
obfuscating, by the computing device using one or more privacy techniques, a portion of the user data; and
transmitting, by the computing device to the server computer system, the obfuscated portion of the user data.
7. The method of claim 6 , wherein the one or more privacy techniques include one or more of the following techniques: differential privacy, homomorphic encryption, and secure multi-party computation.
8. The method of claim 1 , wherein the user request received from the user of the computing device is a request to initiate a transaction, and wherein the decision for the user request is an authorization decision for the transaction.
9. A non-transitory computer-readable medium having instructions stored thereon that are executable by a user computing device to perform operations comprising:
receiving, by a computing device from a server computer system, a baseline model that is generated by the server computer system based on multiple different models trained at different respective user computing devices;
repeatedly retraining, using a stream of user data received at the user computing device, the baseline model to generate a device-trained model, wherein the baseline model is repeatedly retrained at the user computing device without providing user data included in the stream to the server computer system;
transmitting, to the server computer system, the device-trained model and a set of obfuscated user data generated from the stream of user data, wherein the transmitting is performed by the computing device utilizing a first type of wireless communication that provides a higher network performance than a secondary type of wireless communication available to the computing device;
in response to the transmitting, receiving, from the server computer system, an updated version of the device-trained model;
inputting, to the device-trained model, a set of characteristics associated with a user request received from a user of the user computing device, wherein the device-trained model outputs a score for the user request;
transmitting, to the server computer system, the score for the user request, wherein the transmitting includes requesting a decision for the user request; and
in response to receiving a decision for the user request generated by the server computer system based on the set of obfuscated user data, performing an action associated with the user request.
10. The non-transitory computer-readable medium of claim 9 , wherein the user request received from the user of the user computing device is an authentication request to authenticate the user of the user computing device to a user account, and wherein the decision for the user request is an authentication decision.
11. The non-transitory computer-readable medium of claim 9 , wherein the stream of user data includes a plurality of characteristics associated with the user computing device and the set of characteristics associated with the user request.
12. The non-transitory computer-readable medium of claim 9 , wherein performing the action includes automatically granting the user of the user computing device access to an account associated with the server computer system, wherein the automatically granting is performed without requiring credentials from the user.
13. The non-transitory computer-readable medium of claim 10 , wherein the baseline model is an untrained model, and wherein the user computing device generates the device-trained model from scratch.
14. The non-transitory computer-readable medium of claim 10 , wherein the operations further comprise:
obfuscating, using one or more privacy techniques, a portion of the user data; and
transmitting, to the server computer system, the obfuscated portion of the user data.
15. The non-transitory computer-readable medium of claim 10 , wherein the repeatedly training includes:
performing one or more feature engineering techniques on a plurality of characteristics included in the stream of user data, wherein the one or more feature engineering techniques are performed according to one or more conditions of the following conditions: on-the-fly processing, lookup-based processing, and cache-based processing.
16. A user computing system, comprising:
at least one processor; and
a memory having instructions stored thereon that are executable by the at least one processor to cause the user computing system to:
receive, from a server computer system, a baseline model that is generated by the server computer system based on multiple different models;
repeatedly retrain, using a stream of user data received at the user computing system, the baseline model to generate a device-trained model;
transmit, to the server computer system, the device-trained model and a set of obfuscated user data generated at the user computing system from the stream of user data, wherein the transmitting is performed by the user computing system utilizing a first type of wireless communication that provides a higher network performance than a secondary type of wireless communication available to the user computing system;
in response to the transmitting, receive, from the server computer system, an updated version of the device-trained model;
input, to the device-trained model, a set of characteristics associated with an authentication request received from a user of the user computing system, wherein the device-trained model outputs a score for the authentication request;
transmit, to the server computer system, the score for the authentication request, wherein the transmitting includes requesting a decision for the authentication request; and
perform, in response to receiving a decision for the authentication request generated by the server computer system based on the set of obfuscated user data, an action associated with the authentication request.
17. The user computing system of claim 16 , wherein the stream of user data includes a plurality of characteristics associated with the user computing system and the set of characteristics associated with the authentication request, and wherein the set of characteristics includes user account credentials, and wherein the plurality of characteristics include a user device identifier.
18. The user computing system of claim 16 , wherein the instructions are further executable by the at least one processor to cause the user computing system to:
receive, from the server computer system prior to repeatedly training the baseline model, the baseline model, wherein the baseline model is trained by the server computer system without the stream of user data securely stored at the user computing system.
19. The user computing system of claim 16 , wherein the repeatedly retraining includes:
generating an aggregated depiction of data included in the stream of user data; and
comparing different portions of the aggregated depiction with one another, wherein the different portions are selected from different intervals of time within the aggregated depiction; and
adjusting the stream of user data based on one or more portions having data that differ a threshold amount.
20. The user computing system of claim 16 , wherein the instructions are further executable by the at least one processor to cause the user computing system to:
obfuscate, using one or more privacy techniques, a portion of the user data; and transmit, to the server computer system, the obfuscated portion of the user data.