Training models based on balanced training data sets
In some examples, a system balances a number of positive data points and a number of negative data points, to produce a balanced training data set, where the positive data points comprise features associated with authentication events that are positive with respect to an unauthorized classification, and the negative data points comprise features associated with authentication events that are negative with respect to the unauthorized classification. The system trains a plurality of models using the balanced training data set, wherein the plurality of models are trained according to respective different machine learning techniques. The system selects a model from the trained plurality of models based on relative performance of the plurality of models.
1. A non-transitory machine-readable storage medium storing instructions that upon execution cause a system to:
balance a number of positive data points and a number of negative data points, to produce a balanced data set, wherein the positive data points are each indicative of an unauthorized authentication event and each comprise features associated with unauthorized lateral movement in a network, and the negative data points are each indicative of an authorized authentication event and each comprise features associated with the authorized authentication event;
split the balanced data set into a training data set, a test data set, and a validation data set;
train a plurality of models using the training data set, wherein each model of the plurality of models is trained according to respective different machine learning techniques;
evaluate each model of the trained plurality of models using the test data set, wherein evaluating each model of the trained plurality of models comprises evaluating a performance of each model of the plurality of models using the test data set;
select a threshold probability value used by each model of the trained plurality of models to classify an authentication event as positive or negative with respect to the unauthorized lateral movement in the network, wherein the threshold probability value is selected based on the validation training set;
select a model from the trained plurality of models based on the evaluating of each model of the trained plurality of models and a relative performance of each model of the plurality of models; and
apply the selected model to a detected authentication event, wherein applying the model to the detected authentication event distinguishes the detected authentication event as either an authorized authentication event or an unauthorized lateral movement in the network.
2. The non-transitory machine-readable storage medium of claim 1 , wherein the balancing is performed on a collection of positive and negative data points, and the balancing comprises:
sub-sampling the negative data points in the collection of positive and negative data points to reduce the number of the negative data points.
3. The non-transitory machine-readable storage medium of claim 1 , wherein the balancing is performed on a collection of positive and negative data points, and the balancing comprises:
sampling with replacement the positive data points in the collection of positive and negative data points to increase the number of the positive data points.
4. The non-transitory machine-readable storage medium of claim 1 , wherein the balancing is performed on a collection of positive and negative data points, and the balancing comprises:
reducing a difference between the number of positive data points and the number of negative data points.
5. The non-transitory machine-readable storage medium of claim 1 , wherein the plurality of models are trained using two or more of a logistic regression technique, a random forests technique, a gradient boosting technique, or a neural network technique.
6. The non-transitory machine-readable storage medium of claim 1 , wherein selecting a model from the trained plurality of models based on the relative performance of each model of the trained plurality of models is further based on the performance of each model of the trained plurality of models using the validation data set.
7. The non-transitory machine-readable storage medium of claim 1 , wherein the instructions upon execution cause the system to:
apply the selected model on features extracted for a received authentication event to determine whether the received authentication event is positive with respect to the unauthorized lateral movement in the network.
8. The non-transitory machine-readable storage medium of claim 1 , wherein applying the selected model to the detected authentication event distinguishes the detected authentication event as either an authorized authentication event or an unauthorized lateral movement in the network based on the selected threshold probability value for the selected model.
9. A system comprising:
a processor; and
a non-transitory storage medium storing instructions executable on the processor to:
extract features for authentication events and assign labels to collections of the features, the labels comprising a positive label indicative of one or more unauthorized authentication events and each comprising features associated with unauthorized lateral movement in a network, and a negative label indicative of one or more authorized authentication events and each comprising features associated with the one or more authorized authentication events;
produce a balanced data set by balancing a number of collections of features assigned a positive label and a number of collections of features assigned a negative label;
split the balanced data set into a training data set, a test data set, and a validation data set;
train a plurality of models using the training data set, wherein each model of the plurality of models is trained according to respective different machine learning techniques;
evaluate each model of the plurality of models using the test data set, wherein evaluating each model of the plurality of trained models comprises evaluating a performance of each model of the plurality of models using the test data set;
select a threshold probability value used by each model of the trained plurality of models to classify an authentication event as positive or negative with respect to the unauthorized lateral movement in the network and wherein the threshold probability value is selected based on the validation training set;
select a model from the trained plurality of models based on the evaluating of each model of the trained plurality of models and a relative performance of each model of the plurality of models; and
apply the selected model to a detected authentication event, wherein applying the model to the detected authentication event distinguishes the detected authentication event as either an authorized authentication event or an unauthorized lateral movement in the network.
10. The system of claim 9 , wherein the balancing comprises increasing the number of collections of features assigned the positive label relative to the collections of features assigned the negative label.
11. The system of claim 9 , wherein the balancing comprises decreasing the number of collections of features assigned the negative label relative to the collections of features assigned the positive label.
12. The system of claim 9 , wherein the plurality of models is trained using two or more of a logistic regression technique, a random forests technique, a gradient boosting technique, or a neural network technique.
13. The system of claim 9 , wherein the selecting of the model from the trained plurality of models based on a relative performance of each of the plurality of models is further based on the performance of each model of the plurality of models using the validation data set.
14. The system of claim 9 , wherein applying the selected model to the detected authentication event distinguishes the detected authentication event as either an authorized authentication event or an unauthorized lateral movement in the network based on the selected threshold probability value for the selected model.
15. A method comprising:
balancing, by a system comprising a processor, a number of positive data points and a number of negative data points, to produce a balanced data set, wherein the positive data points are each indicative of an unauthorized authentication event and each comprise features associated with an unauthorized lateral movement in a network, and the negative data points are each indicative of an authorized authentication event and each comprise features associated with the authorized authentication event;
splitting, by the system, the balanced data set into a training data set, a test data set, and a validation data set;
training, by the system, a plurality of models using training data set, wherein each model of the plurality of models is trained according to respective different machine learning techniques;
evaluating, by the system, each model of the trained plurality of models using the test data set, wherein evaluating each model of the trained plurality of models comprises evaluating a performance of each model of the plurality of models using the test data set;
selecting, by the system, a threshold probability value used by each model of the trained plurality of models to classify an authentication event as positive or negative with respect to the unauthorized lateral movement in the network, wherein the threshold probability value is selected based on the validation training set;
selecting, by the system, a model from the trained plurality of models based on the evaluating of each model of the trained plurality of models and a relative performance of each model of the trained plurality of models; and
applying, by the system, the selected model to a detected authentication event, wherein applying the model to the detected authentication event distinguishes the detected authentication event as either an authorized authentication event or an unauthorized lateral movement in the network.
16. The method of claim 15 , wherein the selecting a model from the trained plurality of models is based on evaluation of each model of the trained plurality of models based on the relative performance of each model of the trained plurality of models is further based on the performance of each model of the trained plurality of models using the validation data set.
17. The method of claim 15 , wherein the features of the positive and negative data points comprise features extracted from events, the events comprising at least one selected from among: starting a new process at a device, performing a domain name system (DNS) lookup between devices, a transfer of data between devices, a security event on a device, and a Hypertext Transfer Protocol (HTTP) request event.
18. The method of claim 15 , wherein applying the selected model to the detected authentication event distinguishes the detected authentication event as either an authorized authentication event or an unauthorized lateral movement in the network based on the selected threshold probability value for the selected model.
19. The method of claim 15 , wherein the plurality of models are trained using two or more of a logistic regression technique, a random forests technique, a gradient boosting technique, or a neural network technique.
20. The method of claim 15 , wherein selecting a model from the trained plurality of models based on the relative performance of each model of the trained plurality of models is further based on the performance of each model of the trained plurality of models using the validation data set.