Malicious traffic detection with anomaly detection modeling
An anomaly detection model is trained to detect malicious traffic sessions with a low rate of false positives. A sample feature extractor extracts tokens corresponding to human-readable substrings of incoming unstructured payloads in a traffic session. The tokens are correlated with a list of malicious traffic features and frequent malicious traffic features across the traffic session are aggregated into a feature vector of malicious traffic feature frequencies. An anomaly detection model trained on feature vectors for unstructured malicious traffic samples predicts the traffic session as malicious or unclassified. The anomaly detection model is trained and updated based on its' ongoing false positive rate and malicious traffic features in the list of malicious traffic features that result in a high false positive rate are removed.
1. A method comprising:
generating a first plurality of feature vectors for unstructured payloads of one or more malicious traffic sessions, wherein generating the first plurality of feature vectors is based, at least in part, on a plurality of malicious features for the unstructured payloads;
training an anomaly detection model on the first plurality of feature vectors to detect malicious unstructured payloads as non-anomalous and benign unstructured payloads as anomalous; and
based, at least in part, on a false positive rate of the trained anomaly detection model satisfying a performance criterion, wherein the false positive rate comprises a rate of false positives in classifications of the trained anomaly detection model on a second plurality of feature vectors,
updating the plurality of malicious features; and
deploying the trained anomaly detection model.
2. The method of claim 1 , wherein generating the first plurality of feature vectors comprises,
parsing the unstructured payloads into a plurality of strings to obtain the plurality of malicious features of the one or more malicious traffic sessions;
selecting a subset of the plurality of malicious features as longest common substrings from subsets of the plurality of strings; and
generating the first plurality of feature vectors from the subset of the plurality of malicious features.
3. The method of claim 2 , wherein selecting the subset of the plurality of malicious features as longest common substrings from subsets of the plurality of strings comprises,
associating each of the plurality of malicious features with an attack identifier for a corresponding one of the unstructured payloads; and
identifying longest common substrings among features of the plurality of malicious features for each attack identifier.
4. The method of claim 2 , wherein generating the first plurality of feature vectors comprises indicating frequencies of occurrence of features in the first plurality of feature vectors, wherein the frequencies of occurrence of features comprise frequencies of occurrence of features from the subset of the plurality of malicious features in the one or more malicious traffic sessions.
5. The method of claim 2 further comprising, based on parsing the unstructured payloads to obtain the plurality of strings, determining substrings of the unstructured payloads that satisfy criteria for readability.
6. The method of claim 1 , wherein deploying the trained anomaly detection model comprises configuring the trained anomaly detection model to indicate sessions corresponding to non-anomalous-detected unstructured payloads as malicious and sessions corresponding to anomalous-detected unstructured payloads as benign.
7. A non-transitory machine-readable medium having program code stored thereon, the program code comprising instructions to:
generate a feature vector from first features of unstructured payloads of traffic of a network session, wherein the unstructured payloads of the network session comprise unstructured payloads accumulated during monitoring of traffic of the network session over a window size of unstructured payloads, wherein the instructions to generate the feature vector comprise instructions to,
extract a first plurality of strings of the unstructured payloads according to criteria for readability; and
select a subset of the first plurality of strings as the first features based, at least in part, on frequencies of occurrence of a second plurality of strings in the first plurality of strings, wherein the second plurality of strings comprises second features previously identified as corresponding to known malicious traffic;
input the feature vector into an anomaly detection model, wherein the anomaly detection model was trained on feature vectors generated from unstructured payloads of known malicious traffic; and
based on the anomaly detection model classifying the feature vector as non-anomalous, indicate the traffic of the network session as malicious.
8. The non-transitory machine-readable medium of claim 7 , wherein the frequencies of occurrence of the second plurality of strings in the first plurality of strings comprise frequencies of matches of strings in the second plurality of strings with strings in the first plurality of strings.
9. The non-transitory machine-readable medium of claim 7 , wherein the instructions to extract the first plurality of strings of the unstructured payloads according to criteria for readability comprise instructions to,
extract substrings of the unstructured payloads that satisfy the criteria for readability, wherein the criteria for readability comprise that characters in the substrings are in a list of American Standard Code for Information Interchange characters; and
replace substrings of the unstructured payloads that do not satisfy the criteria with placeholder strings.
10. The non-transitory machine-readable medium of claim 7 , wherein the anomaly detection model comprises a one-class support vector machine.
11. An apparatus comprising:
a processor; and
a non-transitory machine-readable medium having instructions stored thereon that are executable by the processor to cause the apparatus to,
generate a first plurality of feature vectors for unstructured payloads of one or more malicious traffic sessions, wherein the instructions executable by the processor to cause the apparatus to generate the first plurality of feature vectors comprise instructions to identify a plurality of malicious features for the one or more malicious traffic sessions;
train an anomaly detection model on the first plurality of feature vectors to detect malicious unstructured payloads as non-anomalous and benign unstructured payloads as anomalous; and
based, at least in part, on a false positive rate of the trained anomaly detection model satisfying a performance criterion, wherein the false positive rate comprises a rate of false positives in classifications of the trained anomaly detection model on a second plurality of feature vectors,
update the plurality of malicious features; and
deploy the trained anomaly detection model for malicious traffic detection.
12. The apparatus of claim 11 , wherein the instructions executable by the processor to cause the apparatus to generate the first plurality of feature vectors comprises instructions to,
parse the unstructured payloads into a plurality of strings to obtain the plurality of malicious features of the one or more malicious traffic sessions;
select a subset of the plurality of malicious features as longest common substrings from subsets of the plurality of strings; and
generate the first plurality of feature vectors based on the subset of the plurality malicious features and the plurality of strings.
13. The apparatus of claim 12 , wherein the instructions executable by the processor to cause the apparatus to select the subset of plurality of malicious features as the longest common substrings from subsets of the plurality of strings comprise instructions to,
associate each of the plurality of strings with an attack identifier of a corresponding unstructured payload; and
identify longest common substrings among subsets of the plurality of strings corresponding to unstructured payloads of each attack identifier.
14. The apparatus of claim 12 , wherein the instructions executable by the processor to cause the apparatus to generate the first plurality of feature vectors comprise instructions to indicate frequencies of occurrence of features in the first plurality of feature vectors, wherein the frequencies of occurrence of features comprise frequencies of occurrence of features from the subset of the plurality of malicious features in the one or more malicious traffic sessions.
15. The apparatus of claim 12 , wherein the instructions executable by the processor to cause the apparatus to parse the unstructured payloads to obtain the plurality of strings comprise instructions to determine substrings of the unstructured payloads that satisfy criteria for readability.
16. The apparatus of claim 15 , wherein the criteria for readability comprise that characters in the substrings are in a list of American Standard Code for Information Interchange characters.
17. The apparatus of claim 11 , wherein the instructions executable by the processor to cause the apparatus to deploy the trained anomaly detection model comprise instructions to indicate sessions corresponding to non-anomalous unstructured payloads detected by the trained anomaly detection model as malicious and sessions corresponding to anomalous unstructured payloads detected by the trained anomaly detection model as benign.
18. The apparatus of claim 11 , wherein the anomaly detection model comprises a one-class support vector machine.