Identifying high-influence features for model-detected anomalies
An anomaly detection system is disclosed capable of reporting anomalous processes or hosts in a computer network using machine learning models trained using unsupervised training techniques. In embodiments, the system assigns observed processes to a set of process categories based on the file system path of the program executed by the process. The system extracts a feature vector for each process or host from the observation records and applies the machine learning models to the feature vectors to determine an outlier metric each process or host. The processes or hosts with the highest outlier metrics are reported as detected anomalies to be further examined by security analysts. In embodiments, the machine learnings models may be periodically retrained based on new observation records using unsupervised machine learning techniques. Accordingly, the system allows the models to learn from newly observed data without requiring the new data to be manually labeled by humans.
1. A system comprising:
one or more computing devices that implement an anomaly detection system, configured to:
execute an anomaly detection model on a dataset of observation records about processes executed on individual machines to determine (a) an outlier record in the dataset, and (b) an outlier score of the outlier record, wherein the anomaly detection model is trained using one or more machine learning techniques;
randomly generate a plurality of synthetic observation record as comparison records to use to determine an influence of a set of features on the outlier score, wherein the generation retains one or more features from the outlier record in an individual comparison record and modifies one or more features in the individual comparison record;
execute the anomaly detection model on the comparison records randomly generated to obtain a set of comparison outlier score; and
output a model interpretation result of the outlier record that indicates respective influences of the set of features on the outlier score.
2. The system of claim 1 , wherein the set of features include one or more features that are randomly selected.
3. The system of claim 1 , wherein the set of features include one or more features that are selected according to user-specified configuration.
4. The system of claim 1 , wherein the anomaly detection system is configured to:
output a determination that the outlier record is an anomaly, wherein the output includes the outlier score of the outlier record and the model interpretation result of the outlier record.
5. The system of claim 1 , wherein the anomaly detection system is configured to:
train another version of the anomaly detection model, wherein the other version is trained using a training dataset that reduces a dimensionality of the dataset of observation records used by the anomaly detection model, where the dimensionality reduction is performed based on the model interpretation result.
6. The system of claim 1 , wherein individual ones of the observation records represent respective observations of machines.
7. The system of claim 1 , wherein individual ones of the observation records represent respective types of processes observed on an individual machine.
8. The system of claim 1 , wherein:
the outlier record is encoded as a bit vector;
to generate the individual comparison record, the anomaly detection system is configured to randomly change one or more bits of the bit vector.
9. The system of claim 1 , wherein the model interpretation result includes respective influence metrics of individual ones of the features in the set of features.
10. The system of claim 9 , wherein to determine an influence metric of a feature, the anomaly detection system is configured to:
fit a linear model to approximate the outlier score based on a linear combination of respective outlier scores computed for the comparison records.
11. The system of claim 1 , wherein the anomaly detection system is configured to:
select which features of the outlier record to modify to generate the comparison records based on a feature variability metric of the features in the dataset.
12. The system of claim 1 , wherein the anomaly detection system is implemented as part of a cyberattack monitoring system, configured to:
collect datasets of observation records from the machines in a remote network; and
use the anomaly detection system to detect evidence of cyberattacks in the collected datasets.
13. A method comprising:
performing, by one or more computing devices that implement an anomaly detection system:
executing an anomaly detection model on a dataset of observation records about processes executed on individual machines to determine (a) an outlier record in the dataset, and (b) an outlier score of the outlier record, wherein the anomaly detection model is trained using one or more machine learning techniques;
randomly generating a plurality of synthetic observation record as comparison records to use to determine an influence of a set of features on the outlier score, wherein the generation retains one or more features from the outlier record in an individual comparison record and modifies one or more features in the individual comparison record;
executing the anomaly detection model on the comparison records randomly generated to obtain a set of comparison outlier score; and
outputting a model interpretation result of the outlier record that indicates respective influences of the set of features on the outlier score.
14. The method of claim 13 , wherein the set of features include one or more features that are randomly selected.
15. The method of claim 13 , wherein the set of features include one or more features that are selected according to user-specified configuration.
16. The method of claim 13 , further comprising performing, by the anomaly detection system:
training another version of the anomaly detection model, wherein the other version is trained using a training dataset that reduces a dimensionality of the dataset of observation records used by the anomaly detection model, where the dimensionality reduction is performed based on the model interpretation result.
17. The method of claim 13 , wherein:
the outlier record is encoded as a bit vector;
generating the individual comparison record comprises randomly changing one or more bits of the bit vector.
18. The method of claim 13 , further comprising performing, by the anomaly detection system:
including in the model interpretation result respective influence metrics of individual ones of the features in the set of features.
19. The method of claim 18 , further comprising performing, by the anomaly detection system:
fitting a linear model to approximate the outlier score based on a linear combination of respective outlier scores computed for the comparison records; and
determining an influence metric of a feature based on the linear model.
20. The system of claim 13 , further comprising performing, by the anomaly detection system:
selecting which features of the outlier record to modify to generate the comparison records based on a feature variability metric of the features in the dataset.