USER BEHAVIOR ANALYTICS FOR INSIDER THREAT DETECTION
Disclosed in some examples are systems, methods, and machine readable mediums for identifying insider threats by determining file system element activity models that correlate to undesirable behavior and then utilizing the determined model to detect insider threats. Events involving file system elements of a client computing device (e.g., a network endpoint) may be monitored by a file system element monitoring application on the client computing device. The values of these signals are aggregated across all events of the same type that have occurred within a predetermined time window (e.g., an hour) for a particular client computing device. Each time an aggregated signal has a value over the threshold, an anomaly is recorded. Anomaly counts for each signal are then calculated as the aggregate number of anomalies for a particular signal over a second time period, the span of which is determined by the generation of first anomaly to the close of an alert by the network monitor. The anomaly counts for the signals are then weighted and summed to produce a risk score.
1 . A method for detecting electronic threats using machine learning, the method comprising:
using one or more hardware processors:
receiving, over a computing network, a plurality of signals associated with file system element events from a computing device endpoint, wherein each signal is categorized by a signal type that defines a characteristic being described by each file system element event;
aggregating values of the signals over a first predetermined period of time to create aggregated signal data;
applying a first machine learning algorithm to the aggregated signal data to determine a dynamic threshold for each signal type, wherein the first machine learning algorithm is trained using historical anomaly counts labeled with indicators of normal or suspicious activity;
comparing the aggregated signal data to each respective dynamic threshold to identify a plurality of anomalies;
calculating a risk score based on the identified plurality of anomalies, wherein the risk score is determined by weighting the anomalies using weights generated by a second machine learning algorithm, the second machine learning algorithm trained using historical data comprising anomaly counts labelled with an indication of whether the anomaly counts indicate normal or suspicious activity; and
sending the risk score to a second computing device for display on a graphical user interface (GUI).
2 . The method of claim 1 , further comprising:
receiving feedback from a network monitor indicating whether the identified anomalies correspond to actual threats; and
using the feedback to refine the first machine learning algorithm, the second machine learning algorithm, or the first and second machine learning algorithms.
3 . The method of claim 1 , wherein calculating the risk score further comprises multiplying each anomaly count by a corresponding weight to produce weighted anomaly counts and summing the weighted anomaly counts.
4 . The method of claim 1 , wherein receiving the plurality of signals includes receiving signals that describe file transfer events, file access events, and file modification events.
5 . The method of claim 1 , further comprising categorizing the file system element events into groups based on user roles and applying different thresholds for each group.
6 . The method of claim 1 , further comprising generating alerts based on the risk score exceeding a predefined threshold.
7 . The method of claim 1 , wherein calculating the risk score includes incorporating time-based decay to prioritize recent anomalies.
8 . A non-transitory machine-readable medium, storing instructions for detecting electronic threats using machine learning, the instructions, which when executed, cause a machine to perform operations comprising:
receiving, over a computing network, a plurality of signals associated with file system element events from a computing device endpoint, wherein each signal is categorized by a signal type that defines a characteristic being described by each file system element event;
aggregating values of the signals over a first predetermined period of time to create aggregated signal data;
applying a first machine learning algorithm to the aggregated signal data to determine a dynamic threshold for each signal type, wherein the first machine learning algorithm is trained using historical anomaly counts labeled with indicators of normal or suspicious activity;
comparing the aggregated signal data to each respective dynamic threshold to identify a plurality of anomalies;
calculating a risk score based on the identified plurality of anomalies, wherein the risk score is determined by weighting the anomalies using weights generated by a second machine learning algorithm, the second machine learning algorithm trained using historical data comprising anomaly counts labelled with an indication of whether the anomaly counts indicate normal or suspicious activity; and
sending the risk score to a second computing device for display on a graphical user interface (GUI).
9 . The non-transitory machine-readable medium of claim 8 , wherein the operations further comprise:
receiving feedback from a network monitor indicating whether the identified anomalies correspond to actual threats; and
using the feedback to refine the first machine learning algorithm, the second machine learning algorithm, or the first and second machine learning algorithms.
10 . The non-transitory machine-readable medium of claim 8 , wherein the operations of calculating the risk score further comprises multiplying each anomaly count by a corresponding weight to produce weighted anomaly counts and summing the weighted anomaly counts.
11 . The non-transitory machine-readable medium of claim 8 , wherein the operations of receiving the plurality of signals includes receiving signals that describe file transfer events, file access events, and file modification events.
12 . The non-transitory machine-readable medium of claim 8 , wherein the operations further comprise categorizing the file system element events into groups based on user roles and applying different thresholds for each group.
13 . The non-transitory machine-readable medium of claim 8 , wherein the operations further comprise generating alerts based on the risk score exceeding a predefined threshold.
14 . The non-transitory machine-readable medium of claim 8 , wherein the operations of calculating the risk score includes incorporating time-based decay to prioritize recent anomalies.
15 . A computing device for detecting electronic threats using machine learning, the computing device comprising:
a hardware processor;
a memory, the memory storing instructions, which when executed by the hardware processor cause the computing device to perform operations comprising:
receiving, over a computing network, a plurality of signals associated with file system element events from a computing device endpoint, wherein each signal is categorized by a signal type that defines a characteristic being described by each file system element event;
aggregating values of the signals over a first predetermined period of time to create aggregated signal data;
applying a first machine learning algorithm to the aggregated signal data to determine a dynamic threshold for each signal type, wherein the first machine learning algorithm is trained using historical anomaly counts labeled with indicators of normal or suspicious activity;
comparing the aggregated signal data to each respective dynamic threshold to identify a plurality of anomalies;
calculating a risk score based on the identified plurality of anomalies, wherein the risk score is determined by weighting the anomalies using weights generated by a second machine learning algorithm, the second machine learning algorithm trained using historical data comprising anomaly counts labelled with an indication of whether the anomaly counts indicate normal or suspicious activity; and
sending the risk score to a second computing device for display on a graphical user interface (GUI).
16 . The computing device of claim 15 , wherein the operations further comprise:
receiving feedback from a network monitor indicating whether the identified anomalies correspond to actual threats; and
using the feedback to refine the first machine learning algorithm, the second machine learning algorithm, or the first and second machine learning algorithms.
17 . The computing device of claim 15 , wherein the operations of calculating the risk score further comprises multiplying each anomaly count by a corresponding weight to produce weighted anomaly counts and summing the weighted anomaly counts.
18 . The computing device of claim 15 , wherein the operations of receiving the plurality of signals includes receiving signals that describe file transfer events, file access events, and file modification events.
19 . The computing device of claim 15 , wherein the operations further comprise categorizing the file system element events into groups based on user roles and applying different thresholds for each group.
20 . The computing device of claim 15 , wherein the operations further comprise generating alerts based on the risk score exceeding a predefined threshold.