Building a ground truth dataset for a machine learning-based security application
Building a ground truth dataset for a machine learning-based security application. In one embodiment, a method may include identifying a set of network devices to add to a ground truth dataset. The method may also include, for each network device in the set of network devices, identifying a potentially malicious application stored on the network device, analyzing behavior of the potentially malicious application to determine whether the potentially malicious application has behaved maliciously, and if so, adding the network device to the ground truth dataset as an infected device or, if not, adding the network device to the ground truth dataset as a clean device. The method may further include training a machine learning classifier of a security application using the ground truth dataset, making a security action decision using the machine learning classifier, and performing a security action on a computer system based on the security action decision.
1. A computer-implemented method for building a ground truth dataset for a machine learning-based security application, at least a portion of the method being performed by a computing device comprising at least one processor, the method comprising:
identifying a set of network devices to be added to a ground truth dataset;
for each network device in the set of network devices to be added to a ground truth dataset:
identifying a potentially malicious application stored on the network device;
analyzing behavior of the potentially malicious application on the network device to determine whether the potentially malicious application has behaved maliciously on the network device; and
if so, adding the network device to the ground truth dataset as an infected network device or, if not, adding the network device to the ground truth dataset as a clean network device;
training a machine learning classifier of a security application using the ground truth dataset;
making a security action decision using the machine learning classifier; and
performing a security action on a computer system based on the security action decision.
2. The method of claim 1 , wherein the analyzing of the behavior of the potentially malicious application on the network device comprises analyzing data gathered from the network device regarding the behavior of the potentially malicious application on the network device.
3. The method of claim 1 , wherein the analyzing of the behavior of the potentially malicious application on the network device comprises building a behavioral graph of the behavior of the potentially malicious application on the network device.
4. The method of claim 3 , wherein the behavioral graph comprises nodes and edges, the nodes representing files downloaded by the potentially malicious application to the network device and the edges representing the behavior of the downloaded files on the network device.
5. The method of claim 1 , wherein the analyzing of the behavior of the potentially malicious application on the network device further comprises monitoring the behavior of the potentially malicious application on the network device for a predetermined period of time to determine whether the potentially malicious application behaved maliciously on the network device during the predetermined period of time.
6. The method of claim 5 , wherein the potentially malicious application is determined to have behaved maliciously on the network device if, during the predetermined period of time, the potentially malicious application downloads a file known to be malicious to the network device and/or the potentially malicious application downloads a file from a URL known to be malicious to the network device.
7. The method of claim 5 , wherein the potentially malicious application is determined to have behaved maliciously on the network device if, during the predetermined period of time, the potentially malicious application exhibits any of the behavior on the network device that is listed for the potentially malicious application in a behavioral malware detection application.
8. The method of claim 7 , wherein the behavioral malware detection application comprises an application that employs file behavioral analysis logs.
9. The method of claim 1 , wherein the identifying of the potentially malicious application comprises identifying the potentially malicious application in a list of potentially malicious applications.
10. The method of claim 9 , wherein the list of potentially malicious applications comprises one or more of Anti-Virus (AV) telemetry data, IPS telemetry data, file reputation telemetry data, and file downloads telemetry data.
11. A computer-implemented method for building a ground truth dataset for a machine learning-based security application, at least a portion of the method being performed by a computing device comprising at least one processor, the method comprising:
identifying a set of network devices be added to a ground truth dataset;
for each network device in the set of network devices to be added to a ground truth dataset:
identify a potentially malicious application stored on the network device;
analyzing data gathered from the network device regarding behavior of the potentially malicious application on the network device to determine whether the potentially malicious application was ever executed on the network device and behaved maliciously on the network device; and
if so, adding the network device to the ground truth dataset as an infected network device or, if not, adding the network device to the ground truth dataset as a clean network device;
training a machine learning classifier of a security application using the ground truth dataset;
making a security action decision using the machine learning classifier; and
performing a security action on a computer system based on the security action decision.
12. The method of claim 11 , wherein:
the analyzing of the behavior of the potentially malicious application on the network device comprises building a behavioral graph of the behavior of the potentially malicious application on the network device; and
the behavioral graph comprises nodes and edges, the nodes representing a file downloaded by the potentially malicious application to the network device and the edges representing the behavior of the downloaded files on the network device.
13. The method of claim 11 , wherein:
the analyzing of the behavior of the potentially malicious application on the network device further comprises monitoring the behavior of the potentially malicious application on the network device for a predetermined period of time to determine whether the potentially malicious application behaved maliciously on the network device during the predetermined period of time; and
the potentially malicious application is determined to have behaved maliciously on the network device if, during the predetermined period of time, the potentially malicious application:
downloads a file known to be malicious to the network device;
downloads a file from a URL known to be malicious to the network device; and/or
exhibits any of the behavior on the network device that is listed for the potentially malicious application in a behavioral malware detection application.
14. The method of claim 11 , wherein the identifying of the potentially malicious application comprises identifying the potentially malicious application in a list of potentially malicious applications.
15. The method of claim 11 , wherein the identifying of the potentially malicious application comprises identifying the potentially malicious application by identifying potentially malicious behavior of the potentially malicious application on the network device using a network-based malware detection application configured to detect one or more of a network connection to a command and control server, sending spam, data exfiltration attempts, and downloading malware.
16. One or more non-transitory computer-readable media comprising one or more computer-readable instructions that, when executed by one or more processors of one or more computing devices, cause the one or more computing devices to perform a method comprising:
identifying a set of network devices to be added to a ground truth dataset;
for each network device in the set of network devices to be added to a ground truth dataset:
identify a potentially malicious application stored on the network device;
analyzing behavior of the potentially malicious application on the network device to determine whether the potentially malicious application has behaved maliciously on the network device; and
if so, adding the network device to the ground truth dataset as an infected network device or, if not, adding the network device to the ground truth dataset as a clean network device;
training a machine learning classifier of a security application using the ground truth dataset;
making a security action decision using the machine learning classifier; and
performing a security action on a computer system based on the security action decision.
17. The one or more non-transitory computer-readable media of claim 16 , wherein:
the making of the security action decision comprises one or more of making a decision that a website is likely malicious, making a decision that the computer system is infected with an application that is likely malicious, making a decision that an extension application of a browser application of the computer system is likely malicious, and making a decision that an email on the computer system is likely malicious; and
the performing of the security action comprises one or more of blocking the computer system from having access to the website, at least partially blocking the computer system from having access to a network, at least partially blocking the extension application from sending data over a network, and quarantining the email from being accessed on the computer system.
18. The one or more non-transitory computer-readable media of claim 16 , wherein the machine learning classifier comprises one or more of a Support Vector Machine, an Artificial Neural Network, and a Deep Learning Network, a Best-First Decision Tree, a C4.5 Tree, a Repeated Incremental Pruning to Produce Error Reduction (RIPPER), and a Fuzzy Unordered Rule Induction Algorithm (FURIA).
19. The one or more non-transitory computer-readable media of claim 16 , wherein:
the analyzing of the behavior of the potentially malicious application on the network device further comprises monitoring the behavior of the potentially malicious application on the network device for a predetermined period of time to determine whether the potentially malicious application behaved maliciously on the network device during the predetermined period of time; and
the potentially malicious application is determined to have behaved maliciously on the network device if, during the predetermined period of time, the potentially malicious application:
downloads a file known to be malicious to the network device;
downloads a file from a URL known to be malicious to the network device; and/or
exhibits any of the behavior on the network device that is listed for the potentially malicious application in a behavioral malware detection application.
20. The one or more non-transitory computer-readable media of claim 16 , wherein the identifying of the potentially malicious application comprises:
identifying the potentially malicious application in a list of potentially malicious applications; or
identifying the potentially malicious application by identifying potentially malicious behavior of the potentially malicious application on the network device using a network-based malware detection application configured to detect one or more of a network connection to a command and control server, sending spam, data exfiltration attempts, and downloading malware.