Systems and methods for adaptively managing data drift in a classifier
The disclosed computer-implemented method for adaptively managing data drift in a classifier may include (i) receiving, at a computing device, an input sample of digital information having an unknown reputation and (ii) performing a security action that may include (A) identifying the input sample as benign or malicious based on a result obtained by classifying the input sample using a machine learning model trained using activity regularization, (B) calculating an internal activity of the machine learning model occurring during the classifying, (C) calculating an activation entropy of the machine learning model occurring during the classifying, (D) comparing a combination of the internal activity and the activation entropy to a threshold, and (E) when the combination of the internal activity and the activation entropy meets or exceeds the threshold, identifying the result as a low-confidence result. Various other methods, systems, and computer-readable media are also disclosed.
1. A computer-implemented method for adaptively managing data drift in a classifier, at least a portion of the method being performed by a computing device comprising at least one processor, the method comprising:
receiving, at the computing device, an input sample of digital information having an unknown reputation; and
performing, at the computing device, a security action comprising:
identifying the input sample as either benign or malicious based on a result obtained by classifying the input sample using a machine learning model trained using activity regularization;
calculating, based at least in part on a pattern of activations of nodes of the machine learning model occurring during the classifying of the input sample, an internal activity of the machine learning model occurring during the classifying of the input sample, wherein a number of activations of nodes of the machine learning model during the classifying of the input sample corresponds to an internal activity; calculating, based at least in part on a pattern of activations of nodes of the machine learning model occurring during the classifying of the input sample, an activation entropy of the machine learning model occurring during the classifying of the input sample, wherein the number of simultaneous activations of nodes of the machine learning model corresponds to an activation entropy;
comparing a combination of the internal activity and the activation entropy to a first threshold; and
in response to determining that the combination of the internal activity and the activation entropy meets or exceeds the first threshold, identifying the result as a low-confidence result.
2. The computer-implemented method of claim 1 , further comprising sending, from the computing device and responsive to identifying the result as a low-confidence result, a notification to a back-end server requesting a machine learning model update.
3. The computer-implemented method of claim 2 , wherein the machine learning model update reduces overfitting.
4. The computer-implemented method of claim 1 , further comprising: storing the low-confidence result in an event queue; and sending, in response to determining that a number of low-confidence results in the event queue meets or exceeds a second threshold and to a back-end server, a notification indicating data drift.
5. The computer-implemented method of claim 4 , further comprising receiving, responsive to the notification and from the back-end server, an instruction to set a monitoring mode.
6. The computer-implemented method of claim 4 , further comprising receiving, responsive to the notification and from the back-end server, a machine learning model update.
7. The computer-implemented method of claim 1 , further comprising sending, to a back-end server, a notification identifying the result as the low-confidence result.
8. The computer-implemented method of claim 7 , further comprising receiving, responsive to the notification identifying the result as the low-confidence result and from a back-end server, a result override.
9. The computer-implemented method of claim 7 , further comprising receiving, responsive to the notification identifying the result as the low-confidence result and from a back-end server, an acknowledgement that the result is a low-confidence result.
10. The computer-implemented method of claim 1 , further comprising:
sending, responsive to identifying the result as the low-confidence result, the input sample to a back-end server; and
receiving a machine learning model update customized for the computing device, wherein the machine learning model update is based at least in part on the input sample.
11. The computer-implemented method of claim 1 , wherein the machine learning model comprises a random forest classifier.
12. The computer-implemented method of claim 1 , wherein the machine learning model comprises a neural forest classifier.
13. The computer-implemented method of claim 1 , wherein the machine learning model comprises a neural network configured as a classifier.
14. The computer-implemented method of claim 1 , further comprising displaying, on a user display, an indication the result is a low-confidence result.
15. A system for adaptively managing data drift in a classifier, the system comprising: a computing device comprising at least one physical processor; and physical memory comprising computer-executable instructions that, when executed by the physical processor, cause the physical processor to: receive, at the computing device, an input sample of digital information having an unknown reputation; and perform, at the computing device, a security action comprising: identifying the input sample as either benign or malicious based on a result obtained by classifying the input sample using a machine learning model trained using activity regularization; calculating, based at least in part on a pattern of activations of nodes of the machine learning model occurring during the classifying of the input sample, an internal activity of the machine learning model occurring during the classifying of the input sample, wherein a number of activations of nodes of the machine learning model during the classifying of the input sample corresponds to an internal activity; calculating, based at least in part on a pattern of activations of nodes of the machine learning model occurring during the classifying of the input sample, an activation entropy of the machine learning model occurring during the classifying of the input sample, wherein the number of simultaneous activations of nodes of the machine learning model corresponds to an activation entropy; comparing a combination of the internal activity and the activation entropy to a first threshold; and in response to determining that the combination of the internal activity and the activation entropy meets or exceeds the first threshold, identifying the result as a low-confidence result.
16. The system of claim 15 , wherein the security action further comprises sending, from the computing device and responsive to identifying the result as a low-confidence result, a notification to a back-end server requesting a machine learning model update.
17. A non-transitory computer-readable medium comprising one or more computer-executable instructions that is configured for execution by at least one processor of a computing device, cause the computing device to: receive, at the computing device, an input sample of digital information having an unknown reputation; and perform, at the computing device, a security action comprising: identifying the input sample as either benign or malicious based on a result obtained by classifying the input sample using a machine learning model trained using activity regularization; calculating, based at least in part on a pattern of activations of nodes of the machine learning model occurring during the classifying of the input sample an internal activity of the machine learning model occurring during the classifying of the input sample, wherein a number of activations of nodes of the machine learning model during the classifying of the input sample corresponds to an internal activity; calculating, based at least in part on a pattern of activations of nodes of the machine learning model occurring during the classifying of the input sample an activation entropy of the machine learning model occurring during the classifying of the input sample, wherein the number of simultaneous activations of nodes of the machine learning model corresponds to an activation entropy; comparing a combination of the internal activity and the activation entropy to a first threshold; and in response to determining that the combination of the internal activity and the activation entropy meets or exceeds the first threshold, identifying the result as a low-confidence result.
18. The non-transitory computer-readable medium of claim 17 , wherein the security action further comprises sending, from the computing device and responsive to identifying the result as a low-confidence result, a notification to a back-end server requesting a machine learning model update.