Multistage analysis of emails to identify security threats
Conventional email filtering services are not suitable for recognizing sophisticated malicious emails, and therefore may allow sophisticated malicious emails to reach inboxes by mistake. Introduced here are threat detection platforms designed to take an integrative approach to detecting security threats. For example, after receiving input indicative of an approval from an individual to access past email received by employees of an enterprise, a threat detection platform can download past emails to build a machine learning (ML) model that understands the norms of communication with internal contacts (e.g., other employees) and/or external contacts (e.g., vendors). By applying the ML model to incoming email, the threat detection platform can identify security threats in real time in a targeted manner.
1. A non-transitory computer-readable medium with instructions stored thereon that, when executed by a processor, cause the processor to perform operations comprising:
receiving an email addressed to a recipient who is an employee of an enterprise;
determining an identity of a sender based on an analysis of the email;
obtaining a first set of models that includes (i) a first model trained using past emails sent by the sender that have been verified as non-malicious and (ii) a second model trained using past emails received by the recipient that have been verified as non-malicious, wherein the first model is trained to identify deviations in features, content, or context that are representative of changes in behavior of the sender, and wherein the second model is trained to identify deviations in features, content, or context that are representative of changes in behavior of the recipient;
applying the first set of models to the email to produce a first output indicative of whether the email is representative of a malicious email;
determining, based on the first output, whether the email is a malicious email;
in response to a determination that the email is a malicious email, applying a second set of models to the email to produce a second output indicative of whether the email is representative of a given type of malicious email; and
performing an action with respect to the email based on the second output.
2. The non-transitory computer-readable medium of claim 1 , wherein the second output indicates that the email is not of the given type of malicious email, and wherein performing the action comprises forwarding the email to an inbox of the employee.
3. The non-transitory computer-readable medium of claim 1 , wherein each model of the second set of models is associated with a different type of malicious email.
4. The non-transitory computer-readable medium of claim 3 , wherein each model in the second set of models produces a separate output when applied to the email, and wherein the operations further comprise applying a third model designed to aggregate the separate outputs produced by the second set of models into a comprehensible visualization component.
5. The non-transitory computer-readable medium of claim 1 , wherein the second output indicates that the email includes a link to a Hypertext Markup Language (HTML) resource, and wherein performing the action comprises:
following the link so that the HTML resource is accessed using a virtual web browser,
extracting a Document Object Model (DOM) for the HTML resource through the virtual web browser, and
analyzing the DOM to determine whether the link represents a security threat.
6. The non-transitory computer-readable medium of claim 1 ,
wherein the second output indicates that the email includes a primary link to a resource hosted by a network-accessible hosting service, and wherein performing the action comprises:
following the primary link so that the resource is accessed using a virtual web browser,
discovering whether any secondary links to secondary resources are present by examining content of the resource through the virtual web browser,
for each secondary link, following the secondary link so that the corresponding secondary resource is accessed using the virtual web browser, and analyzing content of the corresponding secondary resource to determine whether the secondary link represents a security threat, and
determining whether the primary link represents a security threat based on whether any secondary links were determined to represent security threats.
7. The non-transitory computer-readable medium of claim 1 , wherein the second output indicates that the email includes a link to a Hypertext Markup Language (HTML) resource, and wherein performing the action comprises:
following the link so that the HTML resource is accessed using a virtual web browser,
capturing a screenshot of the HTML resource through the virtual web browser,
applying a computer vision algorithm designed to identify similarities between the screenshot and a library of verified sign-in websites, and
determining whether the link represents a security threat based on an output produced by the computer vision algorithm.
8. The non-transitory computer-readable medium of claim 1 , wherein the second output indicates that the email includes an attachment, and wherein performing the action comprises:
opening the attachment within a secure processing environment, and
determining whether the attachment represents a security threat based on an analysis of content of the attachment.
9. A system, comprising:
a processor configured to:
receive an email addressed to a recipient who is an employee of an enterprise;
determine an identity of a sender based on an analysis of the email;
obtain a first set of models that includes: (i) a first model trained using past emails sent by the sender that have been verified as non-malicious, and (ii) a second model trained using past emails received by the recipient that have been verified as non-malicious, wherein the first model is trained to identify deviations in features, content, or context that are representative of changes in behavior of the sender, and wherein the second model is trained to identify deviations in features, content, or context that are representative of changes in behavior of the recipient;
apply the first set of models to the email to produce a first output indicative of whether the email is representative of a malicious email;
determine, based on the first output, whether the email is a malicious email;
in response to a determination that the email is a malicious email, apply a second set of models to the email to produce a second output indicative of whether the email is representative of a given type of malicious email; and
perform an action with respect to the email based on the second output; and
a memory coupled to the processor and configured to provide the processor with instructions.
10. The system of claim 9 , wherein the second output indicates that the email is not of the given type of malicious email, and wherein performing the action comprises forwarding the email to an inbox of the employee.
11. The system of claim 9 , wherein each model of the second set of models is associated with a different type of malicious email.
12. The system of claim 11 , wherein each model in the second set of models produces a separate output when applied to the email, and wherein the operations further comprise applying a third model designed to aggregate the separate outputs produced by the second set of models into a comprehensible visualization component.
13. The system of claim 9 , wherein the second output indicates that the email includes a link to a Hypertext Markup Language (HTML) resource, and wherein performing the action comprises:
following the link so that the HTML resource is accessed using a virtual web browser;
extracting a Document Object Model (DOM) for the HTML resource through the virtual web browser; and
analyzing the DOM to determine whether the link represents a security threat.
14. The system of claim 9 , wherein the second output indicates that the email includes a primary link to a resource hosted by a network-accessible hosting service, and wherein performing the action comprises:
following the primary link so that the resource is accessed using a virtual web browser;
discovering whether any secondary links to secondary resources are present by examining content of the resource through the virtual web browser;
for each secondary link, following the secondary link so that the corresponding secondary resource is accessed using the virtual web browser, and analyzing content of the corresponding secondary resource to determine whether the secondary link represents a security threat; and
determining whether the primary link represents a security threat based on whether any secondary links were determined to represent security threats.
15. The system of claim 9 , wherein the second output indicates that the email includes a link to a Hypertext Markup Language (HTML) resource, and wherein performing the action comprises:
following the link so that the HTML resource is accessed using a virtual web browser;
capturing a screenshot of the HTML resource through the virtual web browser;
applying a computer vision algorithm designed to identify similarities between the screenshot and a library of verified sign-in websites; and
determining whether the link represents a security threat based on an output produced by the computer vision algorithm.
16. The system of claim 9 , wherein the second output indicates that the email includes an attachment, and wherein performing the action comprises:
opening the attachment within a secure processing environment; and
determining whether the attachment represents a security threat based on an analysis of content of the attachment.
17. A method, comprising:
receiving an email addressed to a recipient who is an employee of an enterprise;
determining an identity of a sender based on an analysis of the email;
obtaining a first set of models that includes: (i) a first model trained using past emails sent by the sender that have been verified as non-malicious, and (ii) a second model trained using past emails received by the recipient that have been verified as non-malicious, wherein the first model is trained to identify deviations in features, content, or context that are representative of changes in behavior of the sender, and wherein the second model is trained to identify deviations in features, content, or context that are representative of changes in behavior of the recipient;
applying the first set of models to the email to produce a first output indicative of whether the email is representative of a malicious email;
determining, based on the first output, whether the email is a malicious email;
in response to a determination that the email is a malicious email, applying a second set of models to the email to produce a second output indicative of whether the email is representative of a given type of malicious email; and
performing an action with respect to the email based on the second output.
18. The method of claim 17 , wherein the second output indicates that the email is not of the given type of malicious email, and wherein performing the action comprises forwarding the email to an inbox of the employee.
19. The method of claim 17 , wherein each model of the second set of models is associated with a different type of malicious email.
20. The method of claim 19 , wherein each model in the second set of models produces a separate output when applied to the email, and wherein the operations further comprise:
applying a third model designed to aggregate the separate outputs produced by the second set of models into a comprehensible visualization component.
21. The method of claim 17 , wherein the second output indicates that the email includes a link to a Hypertext Markup Language (HTML) resource, and wherein performing the action comprises:
following the link so that the HTML resource is accessed using a virtual web browser;
extracting a Document Object Model (DOM) for the HTML resource through the virtual web browser; and
analyzing the DOM to determine whether the link represents a security threat.
22. The method of claim 17 , wherein the second output indicates that the email includes a primary link to a resource hosted by a network-accessible hosting service, and wherein performing the action comprises:
following the primary link so that the resource is accessed using a virtual web browser;
discovering whether any secondary links to secondary resources are present by examining content of the resource through the virtual web browser;
for each secondary link, following the secondary link so that the corresponding secondary resource is accessed using the virtual web browser, and analyzing content of the corresponding secondary resource to determine whether the secondary link represents a security threat; and
determining whether the primary link represents a security threat based on whether any secondary links were determined to represent security threats.
23. The method of claim 17 , wherein the second output indicates that the email includes a link to a Hypertext Markup Language (HTML) resource, and wherein performing the action comprises:
following the link so that the HTML resource is accessed using a virtual web browser;
capturing a screenshot of the HTML resource through the virtual web browser;
applying a computer vision algorithm designed to identify similarities between the screenshot and a library of verified sign-in websites; and
determining whether the link represents a security threat based on an output produced by the computer vision algorithm.
24. The method of claim 17 , wherein the second output indicates that the email includes an attachment, and wherein performing the action comprises:
opening the attachment within a secure processing environment; and
determining whether the attachment represents a security threat based on an analysis of content of the attachment.