Securing compromised network devices in a network
Securing compromised network devices in a network. In one embodiment, a method may include (a) identifying a Positive Unlabeled (PU) machine learning classifier, (b) selecting labeled positive samples and unlabeled positive and negative samples as a bootstrap subset of training data from a set of training data, (c) training the PU machine learning classifier, (d) repeating (a)-(c) one or more times to create a set of trained PU machine learning classifiers, (e) predicting probabilities that a network device in a network has been compromised using each of the trained PU machine learning classifiers, (f) combining the probabilities predicted at (e) to generate a combined risk score for the network device, (g) repeating (e)-(f) one or more times to create a ranked list of combined risk scores, and (h) performing a security action on one or more of the network devices in the ranked list.
1. A computer-implemented method for securing compromised network devices in a network, at least a portion of the method being performed by a computing device comprising at least one processor, the method comprising:
(a) identifying a Positive Unlabeled (PU) machine learning classifier;
(b) selecting labeled positive samples and unlabeled positive and negative samples as a bootstrap subset of training data from a set of training data;
(c) training the PU machine learning classifier with the bootstrap subset of training data;
(d) repeating (a)-(c) one or more times to create a set of trained PU machine learning classifiers;
(e) predicting probabilities that a network device in a network has been compromised using each of the trained PU machine learning classifiers in the set of trained PU machine learning classifiers;
(f) combining the probabilities predicted at (e) to generate a combined risk score for the network device;
(g) repeating (e)-(f) one or more times to create a ranked list of combined risk scores; and
(h) performing a security action on one or more of the network devices in the ranked list.
2. The method of claim 1 , wherein a type of the PU machine learning classifier identified in a first iteration of (a) is different from a type of the PU machine learning classifier identified in a second iteration of (a).
3. The method of claim 1 , wherein the PU machine learning classifier identified at (a) comprises one or more of a logistic regression classifier, a Linear and Quadratic Discriminant Analysis (LDA/QDA) classifier, a Support Vector Machine (SVM) classifier, a decision tree classifier, a nearest neighbor classifier, a Naïve Bayes classifier, a neural network classifier, a deep neural network classifier, a deep belief network classifier, a random forest classifier, and a gradient boosted trees classifier.
4. The method of claim 1 , wherein the PU machine learning classifier identified at (a) comprises one or more of a Support Vector Machine (SVM) classifier and a random forest classifier.
5. The method of claim 1 , wherein the predicting of the probabilities that the network device in the network has been compromised comprises predicting probabilities that the network device in the network has been infected with a malicious application.
6. The method of claim 5 , wherein the malicious application comprises one or more of a spyware, a virus, a worm, a logic bomb, a trapdoor, a Trojan horse, a Remote Admin Trojan (RAT), a malware, a mobile malicious code, a malicious font, and a rootkit.
7. The method of claim 1 , wherein each iteration of (b) comprises selecting a different bootstrap subset of training data from the same set of training data.
8. The method of claim 1 , wherein (h) comprises performing the security action on one or more of the network devices positioned highest in the ranked list.
9. A computer-implemented method for securing compromised network devices in a network, at least a portion of the method being performed by a computing device comprising at least one processor, the method comprising:
(a) identifying a Positive Unlabeled (PU) machine learning classifier;
(b) selecting labeled positive samples and unlabeled positive and negative samples as a bootstrap subset of training data from a set of training data, the set of training data comprising fewer labeled positive samples than unlabeled positive and negative samples;
(c) training the PU machine learning classifier with the bootstrap subset of training data;
(d) repeating (a)-(c) one or more times to create a set of trained PU machine learning classifiers, with each iteration of (b) comprising selecting a different bootstrap subset of training data from the same set of training data;
(e) predicting probabilities that a network device in a network has been compromised using each of the trained PU machine learning classifiers in the set of trained PU machine learning classifiers;
(f) combining the probabilities predicted at (e) to generate a combined risk score for the network device;
(g) repeating (e)-(f) one or more times to create a ranked list of combined risk scores; and
(h) performing a security action on one or more of the network devices in the ranked list.
10. The method of claim 9 , wherein (e) comprises using data indicating a security posture of the network device in a network as input to each of the trained PU machine learning classifiers in the set of trained PU machine learning classifiers.
11. The method of claim 10 , wherein the data indicating the security posture of the network device comprises one or more of summary feature data, indicator feature data, temporal feature data, and relational feature data.
12. The method of claim 9 , wherein (f) comprises combining the probabilities predicted at (e) by averaging the probabilities as the combined risk score for the network device.
13. The method of claim 9 , wherein (f) comprises combining the probabilities predicted at (e) by taking the minimum or maximum of the probabilities as the combined risk score for the network device.
14. The method of claim 9 , wherein (f) comprises combining the probabilities predicted at (e) by assigning each of the probabilities to a category and identifying the most frequent category as the combined risk score for the network device.
15. 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:
(a) identifying a Positive Unlabeled (PU) machine learning classifier;
(b) selecting labeled positive samples and unlabeled positive and negative samples as a bootstrap subset of training data from a set of training data;
(c) training the PU machine learning classifier with the bootstrap subset of training data;
(d) repeating (a)-(c) one or more times to create a set of trained PU machine learning classifiers;
(e) predicting probabilities that a network device in a network has been compromised using each of the trained PU machine learning classifiers in the set of trained PU machine learning classifiers;
(f) combining the probabilities predicted at (e) to generate a combined risk score for the network device;
(g) repeating (e)-(f) one or more times to create a ranked list of combined risk scores; and
(h) performing a security action on one or more of the network devices in the ranked list.
16. The one or more non-transitory computer-readable media of claim 15 , wherein:
the labeled positive samples in the set of training data are collected from network devices that have been confirmed to have been compromised; and
the unlabeled positive and negative samples in the set of training data are collected from network devices that are suspected to have been compromised but have not been confirmed to have been compromised.
17. The one or more non-transitory computer-readable media of claim 15 , wherein (f) further comprises calibrating the combined risk score into a scale for display to a user.
18. The one or more non-transitory computer-readable media of claim 15 , wherein the method is performed in a User Behavior Analytics (UBA) application.
19. The one or more non-transitory computer-readable media of claim 15 , wherein the method is performed in a Security Information and Event Management (SIEM) application.
20. The one or more non-transitory computer-readable media of claim 15 , wherein (g) further comprises presenting the ranked list of combined risk scores to a user.