Training a model to detect malicious command and control cloud
The technology disclosed relates to a method, system, and non-transitory computer-readable media that trains a cloud traffic classifier to classify cross-application communications as malicious command and control (C2) traffic or benign cloud traffic. The training uses blocks of malicious Hypertext Transfer Protocol (HTTP) transactions targeted at a plurality of cloud applications by a plurality of clients prequalified as malicious command and control (C2) cloud traffic, and also blocks of benign HTTP transactions targeted at the plurality of cloud applications by the plurality of clients prequalified as benign cloud traffic. A cloud traffic classifier is trained on the cross-application malicious training example set and on the cross-application benign training example set by processing the blocks of the malicious and benign HTTP transactions as inputs, and generating outputs that classify the training examples as respectively malicious C2 cloud traffic or benign cloud traffic.
1. A system including one or more processors coupled to memory, the memory loaded with computer instructions to training a cloud traffic classifier to detect malicious command and control traffic, the instructions, when executed on the processors, implement actions comprising:
memory storing a cross-application malicious training example set, respective malicious training examples in the cross-application malicious training example set including respective blocks of malicious transactions targeted at a plurality of cloud applications by a plurality of clients, wherein the respective blocks of malicious transactions are prequalified as malicious command and control (C2) cloud traffic;
the memory storing a cross-application benign training example set, respective benign training examples in the cross-application benign training example set including respective blocks of benign transactions targeted at a plurality of sanctioned cloud applications by the plurality of clients, wherein the respective blocks of benign transactions are prequalified as benign cloud traffic; and
training logic, having access to the memory, and configured to:
train a cloud traffic classifier on the cross-application malicious training example set to process the respective blocks of malicious transactions as inputs, and generate outputs that classify the respective malicious training examples as the malicious C2 cloud traffic, and
train the cloud traffic classifier on the cross-application benign training example set to process the respective blocks of benign transactions as inputs, and generate outputs that classify the respective benign training examples as the benign cloud traffic,
wherein the trained cloud traffic classifier is configured to detect a beaconing behavior of some transactions in the respective blocks of malicious transactions, wherein the beaconing behavior is detected based on a count of frequency of access by the some transactions to a same URL.
2. The system of claim 1 , wherein the malicious transactions are malicious HyperText Transfer Protocol (HTTP) transactions, and the benign transactions are benign HTTP transactions.
3. The system of claim 1 , wherein the plurality of sanctioned cloud applications is sanctioned by at least one organization.
4. The system of claim 1 , wherein the plurality of clients is managed by at least one organization.
5. The system of claim 1 , wherein the trained cloud traffic classifier is further configured to detect one or more of:
some transactions in the respective blocks of malicious transactions that are en route to an anomalous entity on a cloud application in the plurality of cloud applications, wherein:
an aggregate usage frequency of a particular instance is a measure of how often the particular instance is used by the plurality of clients; and
the particular instance is identified as the anomalous entity based on a determination that the aggregate usage frequency of the particular instance is less than other instances in a plurality of instances;
some transactions in the respective blocks of malicious transactions that originate from an anomalous agent, wherein:
the anomalous agent is a particular web browser in a plurality of web browsers installed on a client of the plurality of clients; and
the particular web browser is identified as the anomalous agent based on a determination that the particular web browser is an obsolete web browser;
some transactions in the respective blocks of malicious transactions that use an anomalous username to access a sanctioned cloud application in the plurality of sanctioned cloud applications, wherein a particular username is identified as an anomalous username based on a determination that the particular username is a previously unused username;
some transactions in the respective blocks of malicious transactions that use an anomalous authentication method to access a sanctioned cloud application in the plurality of sanctioned cloud applications, wherein a particular authentication method is identified as the anomalous authentication method based on a determination that the particular authentication method is a previously unused authentication method.
6. The system of claim 1 , wherein the trained cloud traffic classifier is configured to detect a cat's-paw behavior of some clients in the plurality of clients, wherein the cat's-paw behavior is detected based on a determination that the some clients are repeatedly downloading, deleting, and uploading content from malicious C2 cloud resources.
7. The system of claim 1 , wherein the trained cloud traffic classifier is further configured to detect that some transactions in the respective blocks of malicious transactions are exhibiting an anomalous hostname access pattern behavior, wherein
the anomalous hostname access pattern behavior is detected based on a determination that an unsanctioned executable impersonating as a sanctioned executable of the plurality of sanctioned cloud applications is querying to fewer domain names than the sanctioned executable is configured to use to access the plurality of sanctioned cloud applications.
8. The system of claim 1 , wherein the trained cloud traffic classifier is further configured to detect that some transactions in the respective blocks of malicious transactions are exhibiting an anomalous hostname access pattern behavior, wherein the anomalous hostname access pattern behavior is detected based on an anomalous distribution of queries to different domain names.
9. The system of claim 1 , wherein the trained cloud traffic classifier is further configured to detect that some transactions in the respective blocks of malicious transactions are attempting a malicious task sequence execution, wherein the malicious task sequence execution is detected based on a determination that the some transactions are directed to endpoints hardcoded into a malware detected on a client of the plurality of clients.
10. A computer-implemented method of training a cloud traffic classifier to detect malicious command and control traffic, comprising:
storing a cross-application malicious training example set, respective malicious training examples in the cross-application malicious training example set including respective blocks of malicious transactions targeted at a plurality of cloud applications by a plurality of clients, wherein the respective blocks of malicious transactions are prequalified as malicious command and control (C2) cloud traffic;
storing a cross-application benign training example set, respective benign training examples in the cross-application benign training example set including respective blocks of benign transactions targeted at a plurality of sanctioned cloud applications by the plurality of clients, wherein the respective blocks of benign transactions are prequalified as benign cloud traffic;
training a cloud traffic classifier on the cross-application malicious training example set to process the respective blocks of malicious transactions as inputs, and generate outputs that classify the respective malicious training examples as the malicious C2 cloud traffic; and
training the cloud traffic classifier on the cross-application benign training example set to process the respective blocks of benign transactions as inputs, and generate outputs that classify the respective benign training examples as the benign cloud traffic,
wherein the trained cloud traffic classifier is configured to detect a beaconing behavior of some transactions in the respective blocks of malicious transactions, wherein the beaconing behavior is detected based on a count of frequency of access by the some transactions to a same URL.
11. The computer-implemented method of claim 10 , wherein the trained cloud traffic classifier is further configured to detect one or more of:
some transactions in the respective blocks of malicious transactions that are en route to an anomalous entity on a cloud application in the plurality of cloud applications, wherein:
an aggregate usage frequency of a particular instance is a measure of how often the particular instance is used by the plurality of clients; and
the particular instance is identified as the anomalous entity based on a determination that the aggregate usage frequency of the particular instance is less than other instances in a plurality of instances;
some transactions in the respective blocks of malicious transactions that originate from an anomalous agent, wherein:
the anomalous agent is a particular web browser in a plurality of web browsers installed on a client of the plurality of clients; and
the particular web browser is identified as the anomalous agent based on a determination that the particular web browser is an obsolete web browser;
some transactions in the respective blocks of malicious transactions that use an anomalous username to access a sanctioned cloud application in the plurality of sanctioned cloud applications, wherein a particular username is identified as an anomalous username based on a determination that the particular username is a previously unused username; and
some transactions in the respective blocks of malicious transactions that use an anomalous authentication method to access a sanctioned cloud application in the plurality of sanctioned cloud applications, wherein a particular authentication method is identified as the anomalous authentication method based on a determination that the particular authentication method is a previously unused authentication method.
12. The computer-implemented method of claim 10 , wherein the trained cloud traffic classifier is configured to detect a cat's-paw behavior of some clients in the plurality of clients, wherein the cat's-paw behavior is detected based on a determination that the some clients are repeatedly downloading, deleting, and uploading content from malicious C2 cloud resources.
13. The computer-implemented method of claim 10 , wherein the trained cloud traffic classifier is further configured to detect that some transactions in the respective blocks of malicious transactions are exhibiting an anomalous hostname access pattern behavior, wherein
the anomalous hostname access pattern behavior is detected based on a determination that an unsanctioned executable impersonating as a sanctioned executable of the plurality of sanctioned cloud applications is querying to fewer domain names than a number of domains names the sanctioned executable is configured to use to access the plurality of sanctioned cloud applications.
14. The computer-implemented method of claim 10 , wherein the trained cloud traffic classifier is further configured to detect that some transactions in the respective blocks of malicious transactions are exhibiting an anomalous hostname access pattern behavior, wherein the anomalous hostname access pattern behavior is detected based on an anomalous distribution of queries to different domain names.
15. The computer-implemented method of claim 10 , wherein the trained cloud traffic classifier is further configured to detect that some transactions in the respective blocks of malicious transactions are attempting a malicious task sequence execution, wherein the malicious task sequence execution is detected based on a determination that the some transactions are directed to endpoints hardcoded into a malware detected on a client of the plurality of clients.
16. A non-transitory computer readable storage medium impressed with computer program instructions to training a cloud traffic classifier to detect malicious command and control traffic, the instructions, when executed on a processor, implement a method comprising:
storing a cross-application malicious training example set, respective malicious training examples in the cross-application malicious training example set including respective blocks of malicious transactions targeted at a plurality of cloud applications by a plurality of clients, wherein the respective blocks of malicious transactions are prequalified as malicious command and control (C2) cloud traffic;
storing a cross-application benign training example set, respective benign training examples in the cross-application benign training example set including respective blocks of benign transactions targeted at a plurality of sanctioned cloud applications by the plurality of clients, wherein the respective blocks of benign transactions are prequalified as benign cloud traffic;
training a cloud traffic classifier on the cross-application malicious training example set to process the respective blocks of malicious transactions as inputs, and generate outputs that classify the respective malicious training examples as the malicious C2 cloud traffic; and
training the cloud traffic classifier on the cross-application benign training example set to process the respective blocks of benign transactions as inputs, and generate outputs that classify the respective benign training examples as the benign cloud traffic
wherein the trained cloud traffic classifier is configured to detect a beaconing behavior of some transactions in the respective blocks of malicious transactions, wherein the beaconing behavior is detected based on a count of frequency of access by the some transactions to a same URL.
17. The non-transitory computer readable storage medium of claim 16 , wherein the trained cloud traffic classifier is configured to detect one or more of:
some transactions in the respective blocks of malicious transactions that are en route to an anomalous entity on a cloud application in the plurality of cloud applications, wherein:
an aggregate usage frequency of a particular instance is a measure of how often the particular instance is used by the plurality of clients; and
the particular instance is identified as the anomalous entity based on a determination that the aggregate usage frequency of the particular instance is less than other instances in a plurality of instances;
some transactions in the respective blocks of malicious transactions that originate from an anomalous agent, wherein:
the anomalous agent is a particular web browser in a plurality of web browsers installed on a client of the plurality of clients; and
the particular web browser is identified as the anomalous agent based on a determination that the particular web browser is an obsolete web browser;
some transactions in the respective blocks of malicious transactions that use an anomalous username to access a sanctioned cloud application in the plurality of sanctioned cloud applications,
wherein a particular username is identified as an anomalous username based on a determination that the particular username is a previously unused username;
some transactions in the respective blocks of malicious transactions that use an anomalous authentication method to access a sanctioned cloud application in the plurality of sanctioned cloud applications, wherein a particular authentication method is identified as the anomalous authentication method based on a determination that the particular authentication method is a previously unused authentication method;
a cat's-paw behavior of some clients in the plurality of clients, wherein the cat's-paw behavior is detected based on a determination that the some clients are repeatedly downloading, deleting, and uploading content from malicious C2 cloud resources;
the trained cloud traffic classifier is further configured to detect that some transactions in the respective blocks of malicious transactions are exhibiting an anomalous hostname access pattern behavior,
wherein the anomalous hostname access pattern behavior is detected based on a determination that an unsanctioned executable impersonating as a sanctioned executable of the sanctioned cloud application is querying to fewer domain names than a number of domains names the sanctioned executable is configured to use to access the sanctioned cloud application; and
malicious task sequence execution is detected based on a determination that the some transactions are directed to endpoints hardcoded into a malware detected on a client of the plurality of clients.
18. The non-transitory computer readable storage medium of claim 16 , wherein the malicious transactions are malicious HyperText Transfer Protocol (HTTP) transactions, and the benign transactions are benign HTTP transactions.