One-shot behavioral biometrics for login authentication using machine learning model
In one approach, a method includes: receiving a reference login event input from a user, the reference login event input being associated with a first session of the user logging into an account; receiving a new login event input from the user, the new login event input being associated with a second session of the user logging into the account; accessing a machine learning model, wherein the machine learning model is trained using data selected based on a similarity of behavior between different users; and authenticating, with the machine learning model, the user for the account, based at least in part on the reference login event input and the new login event input. In examples, the reference and new login event inputs comprise one or more items of biometric data generated by interaction of the user in a web environment and/or a mobile environment for logging into the account.
1. A method comprising:
receiving a reference login event input from a new user, the reference login event input being associated with a first session of the new user logging into an account;
receiving a new login event input from the new user, the new login event input being associated with a second session of the new user logging into the account;
accessing a machine learning model, wherein the machine learning model is trained using data selected based on a closest similarity of behavior between different prior users, wherein the prior users do not include the new user; and
authenticating, with the machine learning model, the new user for the account, based at least in part on the reference login event input and the new login event input.
2. The method of claim 1 , wherein the machine learning model is trained with login event input associated with a plurality of users logging into an account on a computer system.
3. The method of claim 1 , wherein the new login event input comprises one or more items of biometric data associated with the new user.
4. The method of claim 3 , wherein an item of the one or more items of biometric data associated with the new user is generated by interaction of the new user with an input device or a handheld device for logging into the account, the interaction communicating a login credential of the new user.
5. The method of claim 4 , wherein the interaction is performed by the new user in a web environment, or a mobile device environment.
6. The method of claim 4 , wherein an item of the one or more items of biometric data associated with the new user is keyboard event biometric data, mouse event biometric data, motion sensor biometric data, touch event biometric data, or hand gesture event biometric data.
7. The method of claim 6 , wherein the motion sensor biometric data is obtained using an inertial sensor.
8. The method of claim 7 , wherein the inertial sensor includes at least one of: a gyroscope sensor, a magnetometer sensor, or an accelerometer sensor.
9. The method of claim 6 , wherein the keyboard event biometric data comprises a set of features associated with keystrokes of the new user during the second session.
10. The method of claim 6 , wherein the mouse event biometric data comprises a set of features associated with dynamics of a mouse movement of the new user during the second session.
11. The method of claim 6 , wherein the touch event biometric data comprises a set of features associated with touches of the new user on a handheld device during the second session.
12. The method of claim 6 , wherein a vector of motion sensor features is fused with a vector of touch features to generate a vector of login features.
13. The method of claim 1 , wherein the machine learning model is trained by a plurality of positive pairs of training data and a plurality of negative pairs of training data.
14. The method of claim 1 , wherein the machine learning model comprises a Siamese neural network trained by a plurality of triplets of training data.
15. The method of claim 13 , wherein a positive pair of the plurality of positive pairs of training data is a hard positive pair of training data, and a negative pair of the plurality of negative pairs of training data is a hard negative pair of training data.
16. The method of claim 13 , wherein the prior users include a first user and a second user, the plurality of negative pairs of training data comprises a closest negative pair of training data selected based on an Euclidean distance representing a degree of similarity between login features of the first user and login features of the second user.
17. The method of claim 16 , wherein the closest negative pair of training data is selected based on a pre-clustered set of training data.
18. The method of claim 14 , wherein the plurality of triplets of training data comprises semi hard triplets of training data.
19. A non-transitory computer-readable medium storing instructions that, when executed, cause a computer system to:
receive a reference login event input from a new user, the reference login event input being associated with a first session of the new user logging into an account;
receive a new login event input from the new user, the new login event input being associated with a second session of the new user logging into the account;
access a machine learning model, wherein the machine learning model is trained using data selected based on a degree of similarity between behavior of a first prior user and behavior of a second prior user; and
authenticate, with the machine learning model, the new user for the account, based at least in part on the reference login event input and the new login event input.
20. A system comprising:
at least one processor; and
memory in communication with the processor and storing instructions that, when executed by the processor, cause the system to:
receive a reference login event input from a new user, the reference login event input being associated with a first session of the new user logging into an account;
receive a new login event input from the new user, the new login event input being associated with a second session of the new user logging into the account;
access a machine learning model, wherein the machine learning model is trained using data selected based on a similarity of behavior between different prior users other than the new user, the selected data comprises a plurality of negative pairs, and each negative pair is selected based on similarity between login features of a respective first prior user and login features of a respective second prior user; and
authenticate, with the machine learning model, the new user for the account, based at least in part on the reference login event input and the new login event input.