US20160182545A1

US20160182545A1 - Methods, systems, and media for masquerade attack detection by monitoring computer user behavior

Info

Publication number: US20160182545A1
Application number: US15/056,627
Authority: US
Inventors: Salvatore J. Stolfo; Malek Ben Salem; Shlomo Hershkop
Original assignee: Columbia University of New York
Current assignee: Columbia University of New York
Priority date: 2008-12-02
Filing date: 2016-02-29
Publication date: 2016-06-23
Also published as: US20100269175A1; US9311476B2; US8769684B2; US20160065614A1

Abstract

Methods, systems, and media for masquerade attack detection by monitoring computer user behavior are provided. In accordance with some embodiments, a method for detecting masquerade attacks is provided, the method comprising: monitoring, using a hardware processor, a first plurality of user actions in a computing environment; generating a user intent model based on the first plurality of user actions; monitoring a second plurality of user actions in the computing environment; determining whether at least one of the second plurality of user actions deviates from the generated user intent model; determining whether the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information in response to determining that at least one of the second plurality of user actions deviates from the generated user intent model; and generating an alert in response to determining that the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information.

Description

CROSS REFERENCE TO RELATED APPLICATIONS

This application is a continuation of U.S. patent application Ser. No. 14/272,099, filed May 7, 2014, which is a continuation of U.S. patent application Ser. No. 12/628,587, filed Dec. 1, 2009, which claims the benefit of U.S. Provisional Patent Application No. 61/119,294, filed Dec. 2, 2008 and U.S. Provisional Application No. 61/119,540, filed Dec. 3, 2008, which are hereby incorporated by reference herein in their entireties.

STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

The invention was made with government support under Grant No. 60NANB1D0127 awarded by the U.S. Department of Homeland Security through the Institute for Information Infrastructure Protection (I3P) and under Grant No. W911NF-06-1-0151-49626-CI awarded by the Army Research Office (ARO). The government has certain rights in the invention.

TECHNICAL FIELD

The disclosed subject matter relates to methods, systems, and media for masquerade attack detection by monitoring computer user behavior.

BACKGROUND

Much research in computer security has focused on approaches for preventing unauthorized and illegitimate access to systems and information. However, one of the most damaging malicious activities is the result of internal misuse within an organization. This may be because much of the attention has been focused on preventative measures against computer viruses, worms, trojans, hackers, rootkits, spyware, key recovery attacks, denial-of-service attacks, malicious software (or malware), probes, etc. such that far less attention has been focused inward.
Insider threats generally include masqueraders and/or traitors. Masqueraders generally include attackers that impersonate another inside user, while traitors generally include inside attackers that use their own legitimate credentials to attain illegitimate goals. For example, identity theft in financial transaction systems is one example of a masquerade attack. Upon stealing a bank customer's commercial identity (e.g., their credit card or account information), a masquerader presents those credentials for the malicious purpose of using the customer's credit line to steal money. In another example, a disgruntled insider employee is an example of a traitor, where the traitor generally has full knowledge of the systems he or she routinely uses and the security policies in force and uses his or her knowledge and access to conduct unauthorized activities. In yet another example, the disgruntled insider employee can act as a traitor and a masquerader upon stealing the identity of another employee within the organization.
In addition, some external attackers can become inside attackers when, for example, an external attacker gains internal network access. For example, external attackers can gain access to an internal network with the use of spyware or rootkits. Such software can be easily installed on computer systems from physical or digital media (e.g., email, downloads, etc.) and can provide an attacker with administrator or “root” access on a machine along with the capability of gathering sensitive data. In particular, the attacker can snoop or eavesdrop on a computer or a network, download and exfiltrate data, steal assets and information, destroy critical assets and information, and/or modify information. Rootkits have the ability to conceal themselves and elude detection, especially when the rootkit is previously unknown, as is the case with zero-day attacks. An external attacker that manages to install a rootkit internally in effect becomes an insider, thereby multiplying the ability to inflict harm.
Current detection approaches generally monitor command line calls issued by users, system calls for unusual application use or events, database or file accesses, and the organization policy management rules and compliance logs. For example, one particular detection approach detects malicious insider activities by specifically monitoring violations of a “Need-to-Know” policy. Another approach builds an adaptive command line interface. However, these approaches failed to reveal or clarify the user's intent when issuing commands or running processes. In addition, these detection approaches produce results with unacceptable false positive rates.
There is therefore a need in the art for approaches for detecting masquerade attacks by monitoring computer user behavior. Accordingly, it is desirable to provide methods, systems and media that overcome these and other deficiencies of the prior art.

SUMMARY

In accordance with various embodiments, mechanisms for detecting masquerade attacks by monitoring computer user behavior are provided.
Insider attacks generally include attacks from traitors and attacks from masqueraders that have already obtained credentials to access a computing environment. A traitor can be a legitimate user within an organization that has been granted access to systems and other information resources, but whose actions are counter to policy and whose goal is to negatively affect confidentiality, integrity, or availability of some information asset. A masquerader can be an attacker who succeeds in stealing a legitimate user's identity and impersonates the legitimate user for malicious purposes. For example, once a masquerader steals a bank customer's commercial identity including credit card and/or account information, the masquerader presents that information for the malicious purpose of using the customer's credit line to steal money. Inside attackers, traitors, and masqueraders are sometimes referred to herein as “masquerader” or “attacker.”
However, the masquerader is generally unlikely to know how the victim computer user behaves when using a file system. For example, each individual computer user generally knows his or her own file system well enough to search in a limited, targeted, and unique fashion in order to find information germane to the current task. In another example, some individual computer users initiate the same or similar commands each time they log in to set their environment before using it—e.g., initiate some set of applications, such as read e-mail, open a web browser, and/or start a chat session. Masqueraders, on the other hand, generally do not know the user's file system and/or the layout of the user's desktop. As such, masqueraders generally search more extensively and broadly in a manner that is different from the victim user being impersonated. For example, a masquerader is likely to engage in extensive search activities since the masquerader may be unfamiliar with the system local resources or the masquerader may be exploring the file system prior to launching an attack.
In some embodiments, masquerade detection mechanisms are provided for monitoring a user's actions and determining whether current user actions are consistent with the user's historical behavior. For example, these mechanisms create one or more models of the behavior and intent of a particular user and determine whether the current user actions deviate from the one or more models. In another example, user actions (e.g., user commands, Windows applications, registry-based activities, DLL activities, etc.) are classified into categories allowing models to more easily detect patterns of behavior that are indicative of user intent. In a more particular example, a behavior related to search activities can be modeled to detect unusual amounts of search activities on a file system, which can be an indication that a masquerader has gained entry to the file system. As such, a search category for the particular user can be modeled.
In some embodiments, these masquerade detection mechanisms can include traps that can be used to confuse, deceive, and/or detect nefarious inside attackers that attempt to exfiltrate and/or use information. These traps use decoy information (sometimes referred to herein as “bait information,” “bait traffic,” “decoy media,” or “decoy documents”) to attract, deceive, and/or confuse attackers. For example, large amounts of decoy information can be generated and inserted into the network flows and large amount of decoy documents, or documents containing decoy information, can be generated and placed within a file system to lure potential attackers. In another example, decoy documents can be generated that are machine-generated documents containing content to entice an inside attacker into stealing bogus information. Among other things, decoy information can be used to reduce the level of system knowledge of an attacker, entice the attacker to perform actions that reveal their presence and/or identities, and uncover and track the unauthorized activities of the attacker.
It should be noted that files containing decoy information (e.g., a decoy password file, a file with bogus credit card information, etc.) can be used to reduce and/or eliminate false positives generated by detectors using the above-mentioned masquerade detection model. If the user accessed a file containing decoy information and performed abnormal behavior prior to or concurrently with accessing the decoy file, a detector or a sensor can use this information to determine that a masquerade attack is likely occurring. The masquerade detection model can be used to reduce and/or eliminate false positives generated by detectors that monitor decoy files.
Accordingly, masquerade detections mechanisms are provided that create and use a model of the computer user's behavior to determine whether at least one of the current user actions is deemed abnormal and that detect whether at least one of the current user actions includes accessing, transmitting, opening, executing, and/or misusing decoy information on the file system.
In some embodiments, a method for detecting masquerade attacks is provided, the method comprising: monitoring a first plurality of user actions and access of decoy information in a computing environment; generating a user intent model for a category that includes at least one of the first plurality of user actions; monitoring a second plurality of user actions; comparing the second plurality of user actions with the user intent model by determining deviation from the generated user intent model; identifying whether the second plurality of user actions is a masquerade attack based at least in part on the comparison; and generating an alert in response to identifying that the second plurality of user actions is the masquerade attack and in response to determining that the second plurality of user actions includes accessing the decoy information in the computing environment.
In some embodiments, the detection is performed using a host-based anomaly sensor. The host-based sensor can, for example, continuously monitor user behavior and detect significant behavior differences or changes, which can be indicative of a security breach or malicious intent. The host-based sensor can also continuously monitor whether a file containing decoy information has been accessed.
In some embodiments, the user actions are classified into contextual categories and models of user intent are generated for one or more categories. These categories can be selected for modeling based on, for example, operating environment, user, etc.
In accordance with some embodiments, a system for detecting masquerade attacks is provided, the system comprising a processor that: monitors a first plurality of user actions and access of decoy information in a computing environment; generates a user intent model for a category that includes at least one of the first plurality of user actions; monitors a second plurality of user actions; compares the second plurality of user actions with the user intent model by determining deviation from the generated user intent model; identifies whether the second plurality of user actions is a masquerade attack based at least in part on the comparison; and generates an alert in response to identifying that the second plurality of user actions is the masquerade attack and in response to determining that the second plurality of user actions includes accessing the decoy information in the computing environment.
In accordance with some embodiments, a non-transitory computer-readable medium containing computer-executable instructions that, when executed by a processor, cause the processor to perform a method for detecting masquerade attacks is provided. The method comprises: monitoring a first plurality of user actions and access of decoy information in a computing environment; generating a user intent model for a category that includes at least one of the first plurality of user actions; monitoring a second plurality of user actions; comparing the second plurality of user actions with the user intent model by determining deviation from the generated user intent model; identifying whether the second plurality of user actions is a masquerade attack based at least in part on the comparison; and generating an alert in response to identifying that the second plurality of user actions is the masquerade attack and in response to determining that the second plurality of user actions includes accessing the decoy information in the computing environment.
In accordance with some embodiments, a method for detecting masquerade attacks is provided, the method comprising: monitoring, using a hardware processor, a first plurality of user actions in a computing environment; generating a user intent model based on the first plurality of user actions; monitoring a second plurality of user actions in the computing environment; determining whether at least one of the second plurality of user actions deviates from the generated user intent model; determining whether the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information in response to determining that at least one of the second plurality of user actions deviates from the generated user intent model; and generating an alert in response to determining that the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information.
In accordance with some embodiments, a system for detecting masquerade attacks is provided, the system comprising: at least one hardware processor that is configured to: monitor a first plurality of user actions in a computing environment; generate a user intent model based on the first plurality of user actions; monitor a second plurality of user actions in the computing environment; determine whether at least one of the second plurality of user actions deviates from the generated user intent model; determine whether the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information in response to determining that at least one of the second plurality of user actions deviates from the generated user intent model; and generate an alert in response to determining that the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information.
In accordance with some embodiments, a non-transitory computer-readable medium containing computer-executable instructions that, when executed by a processor, cause the processor to perform a method for detecting masquerade attacks is provided. The method comprises: monitoring a first plurality of user actions in a computing environment; generating a user intent model based on the first plurality of user actions; monitoring a second plurality of user actions in the computing environment; determining whether at least one of the second plurality of user actions deviates from the generated user intent model; determining whether the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information in response to determining that at least one of the second plurality of user actions deviates from the generated user intent model; and generating an alert in response to determining that the second plurality of user actions include performing an action on a file in the computing environment that contains decoy information.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a diagram of a system suitable for implementing an application that detects masquerade attacks in accordance with some embodiments of the disclosed subject matter.

FIG. 2 is an exemplary system architecture that includes host-based sensors for monitoring user behavior, creating models of user behavior intent, and detect differences or changes in user behavior in accordance with some embodiments of the disclosed subject matter.

FIG. 3 is a diagram showing an example of a registry access sensor monitoring a user computing environment in accordance with some embodiments of the disclosed subject matter.

FIG. 4 is a diagram showing an example of a process for detecting masquerade attacks by monitoring whether current user behavior is abnormal as compared with previous user behavior and monitoring decoy information in accordance with some embodiments of the disclosed subject matter.

FIG. 5 is a diagram showing an example of a process for identifying whether current user behavior is a masquerade attack based on deviations from a generated model of user intent in accordance with some embodiments of the disclosed subject matter.

FIG. 6 is a diagram showing an example of a taxonomy of UNIX commands that are used to abstract sequences of user commands and actions in accordance with some embodiments of the disclosed subject matter.

FIG. 7 is a diagram showing an example of a taxonomy of Microsoft Windows applications that are used to abstract applications, DLL activities, and/or registry-based activities in accordance with some embodiments of the disclosed subject matter.

DETAILED DESCRIPTION

In accordance with various embodiments, mechanisms for detecting masquerade attacks by monitoring computer user behavior are provided.
Insider attacks generally include attacks from traitors and attacks from masqueraders that have already obtained credentials to access a computing environment. A traitor can be a legitimate user within an organization that has been granted access to systems and other information resources, but whose actions are counter to policy and whose goal is to negatively affect confidentiality, integrity, or availability of some information asset. A masquerader can be an attacker who succeeds in stealing a legitimate user's identity and impersonates the legitimate user for malicious purposes. For example, once a masquerader steals a bank customer's commercial identity including credit card and/or account information, the masquerader presents that information for the malicious purpose of using the customer's credit line to steal money. Inside attackers, traitors, and masqueraders are sometimes referred to herein as “masquerader” or “attacker.”
However, the masquerader is generally unlikely to know how the victim computer user behaves when using a file system. For example, each individual computer user generally knows his or her own file system well enough to search in a limited, targeted, and unique fashion in order to find information germane to the current task. In another example, some individual computer users initiate the same or similar commands each time they log in to set their environment before using it—e.g., initiate some set of applications, such as read e-mail, open a web browser, and/or start a chat session. Masqueraders, on the other hand, generally do not know the user's file system and/or the layout of the user's desktop. As such, masqueraders generally search more extensively and broadly in a manner that is different from the victim user being impersonated. For example, a masquerader is likely to engage in extensive search activities since the masquerader may be unfamiliar with the system local resources or the masquerader may be exploring the file system prior to launching an attack.
In some embodiments, masquerade detection mechanisms are provided for monitoring a user's actions and determining whether current user actions are consistent with the user's historical behavior. For example, these mechanisms create one or more models of the behavior and intent of a particular user and determine whether the current user actions deviate from the one or more models. In another example, user actions (e.g., user commands, Windows applications, registry-based activities, DLL activities, etc.) are classified into categories allowing models to more easily detect patterns of behavior that are indicative of user intent. In a more particular example, a behavior related to search activities can be modeled to detect unusual amounts of search activities on a file system, which can be an indication that a masquerader has gained entry to the file system. As such, a search category for the particular user can be modeled.
In some embodiments, these masquerade detection mechanisms can include traps that can be used to confuse, deceive, and/or detect nefarious inside attackers that attempt to exfiltrate and/or use information. These traps use decoy information (sometimes referred to herein as “bait information,” “bait traffic,” “decoy media”, or “decoy documents”) to attract, deceive, and/or confuse attackers. For example, large amounts of decoy information can be generated and inserted into the network flows and large amount of decoy documents, or documents containing decoy information, can be generated and placed within a file system to lure potential attackers. In another example, decoy documents can be generated that are machine-generated documents containing content to entice an inside attacker into stealing bogus information. Among other things, decoy information can be used to reduce the level of system knowledge of an attacker, entice the attacker to perform actions that reveal their presence and/or identities, and uncover and track the unauthorized activities of the attacker.
It should be noted that files containing decoy information (e.g., a decoy password file, a file with bogus credit card information, etc.) can be used to reduce and/or eliminate false positives generated by detectors using the above-mentioned masquerade detection model. If the user accessed a file containing decoy information and performed abnormal behavior prior to or concurrently with accessing the decoy file, a detector or a sensor can use this information to determine that a masquerade attack is likely occurring. The masquerade detection model can be used to reduce and/or eliminate false positives generated by detectors that monitor decoy files.
Accordingly, masquerade detections mechanisms are provided that create and use a model of the computer user's behavior to determine whether at least one of the current user actions is deemed abnormal and that detect whether at least one of the current user actions includes accessing, transmitting, opening, executing, and/or misusing decoy information on the file system.
These mechanisms can be used in a variety of applications. For example, a host agent (e.g., an ActiveX control, a Javascript control, etc.) can insert and monitor decoy password information with an embedded active beacon among data in Microsoft Outlook (e.g., in the address book, in the notes section, etc.), while also generating models of normal user behavior and monitoring current user actions to determine whether a deviation from those models has occurred.
Turning to FIG. 1, an example of a system 100 in which the masquerade attack detection mechanisms can be implemented. As illustrated, system 100 includes multiple collaborating computer systems 102, 104, and 106, a communication network 108, a masquerader/compromised computer 110, communication links 112, detection and deception system 114, and an attacker computer system 116.
Collaborating systems 102, 104, and 106 can be systems owned, operated, and/or used by universities, businesses, governments, non-profit organizations, families, individuals, and/or any other suitable person and/or entity. Collaborating systems 102, 104, and 106 can include any number of user computers, servers, firewalls, routers, switches, gateways, wireless networks, wired networks, intrusion detection systems, and any other suitable devices. Collaborating systems 102, 104, and 106 can include one or more processors, such as a general-purpose computer, a special-purpose computer, a digital processing device, a server, a workstation, and/or various other suitable devices. Collaborating systems 102, 104, and 106 can run programs, such as operating systems (OS), software applications, a library of functions and/or procedures, background daemon processes, and/or various other suitable programs. In some embodiments, collaborating systems 102, 104, and 106 can support one or more virtual machines. Any number (including only one) of collaborating systems 102, 104, and 106 can be present in system 100, and collaborating systems 102, 104, and 106 can be identical or different.
Communication network 108 can be any suitable network for facilitating communication among computers, servers, etc. For example, communication network 108 can include private computer networks, public computer networks (such as the Internet), telephone communication systems, cable television systems, satellite communication systems, wireless communication systems, any other suitable networks or systems, and/or any combination of such networks and/or systems.
Generally speaking, a user of masquerader/compromised computer 110 is a masquerader or an inside attacker that legitimately has access to communications network 108 and/or one or more systems 102, 104, and 106, but uses his or her access to attain illegitimate goals. For example, a user of masquerader/compromised computer 110 can be a traitor that uses his or her own legitimate credentials to gain access to communications network 108 and/or one or more systems 102, 104, and 106, but uses his or her access to attain illegitimate goals. In another example, a user of masquerader/compromised computer 110 can be a masquerader that impersonates another inside user. In addition, masquerader/compromised computer 110 can be any computer, server, or other suitable device for launching a computer threat, such as a virus, worm, trojan, rootkit, spyware, key recovery attack, denial-of-service attack, malware, probe, etc.
It should be noted that, in some embodiments, an external attacker can become an inside attacker when the external attacker obtains internal network access. For example, using spyware or rootkits, external attackers can gain access to communications network 108. Such software can easily be installed on computer systems from physical or digital media (e.g., email, downloads, etc.) that provides an external attacker with administrator or “root” access on a machine along with the capability of gathering sensitive data. The external attacker can also snoop or eavesdrop on one or more systems 102, 104, and 106 or communications network 108, download and exfiltrate data, steal assets and information, destroy critical assets and information, and/or modify information. Rootkits have the ability to conceal themselves and elude detection, especially when the rootkit is previously unknown, as is the case with zero-day attacks. An external attacker that manages to install rootkits internally in effect becomes an insider, thereby multiplying the ability to inflict harm.
In some embodiments, the owner of masquerader/compromised computer 110 may not be aware of what operations masquerader/compromised computer 110 is performing or may not be in control of masquerader/compromised computer 110. Masquerader/compromised computer 110 can be acting under the control of another computer (e.g., attacking computer system 116) or autonomously based upon a previous computer attack which infected computer 110 with a virus, worm, trojan, spyware, malware, probe, etc. For example, some malware can passively collect information that passes through masquerader/compromised computer 110. In another example, some malware can take advantage of trusted relationships between masquerader/compromised computer 110 and other systems 102, 104, and 106 to expand network access by infecting other systems. In yet another example, some malware can communicate with attacking computer system 116 through an exfiltration channel 120 to transmit confidential information (e.g., IP addresses, passwords, credit card numbers, etc.).
It should be noted that malicious code can be injected into an object that appears as an icon in a document. In response to manually selecting the icon, the malicious code can launch an attack against a third-party vulnerable application. Malicious code can also be embedded in a document, where the malicious code does not execute automatically. Rather, the malicious code lies dormant in the file store of the environment awaiting a future attack that extracts the hidden malicious code.
Alternatively, in some embodiments, masquerader/compromised computer 110 and/or attacking computer system 116 can be operated by an individual or organization with nefarious intent. For example, with the use of malicious code and/or exfiltration channel 120, a user of masquerader/compromised computer 110 or a user of attacking computer system 116 can perform can perform unauthorized activities (e.g., exfiltrate data without the use of channel 120, steal information from one of the collaborating systems 102, 104, and 106), etc.
It should be noted that any number of masquerader/compromised computers 110 and attacking computer systems 116 can be present in system 100, but only one is shown in FIG. 1 to avoid overcomplicating the drawing.
More particularly, for example, each of the one or more collaborating or client computers 102, 104, and 106, masquerader/compromised computer 110, detection and deception system 114, and attacking computer system 116, can be any of a general purpose device such as a computer or a special purpose device such as a client, a server, etc. Any of these general or special purpose devices can include any suitable components such as a processor (which can be a microprocessor, digital signal processor, a controller, etc.), memory, communication interfaces, display controllers, input devices, etc. For example, collaborating system 102 can be implemented as a personal computer, a personal data assistant (PDA), a portable email device, a multimedia terminal, a mobile telephone, a set-top box, a television, etc.
In some embodiments, any suitable computer readable media can be used for storing instructions for performing the processes described herein. For example, in some embodiments, computer readable media can be transitory or non-transitory. For example, non-transitory computer readable media can include media such as magnetic media (such as hard disks, floppy disks, etc.), optical media (such as compact discs, digital video discs, Blu-ray discs, etc.), semiconductor media (such as flash memory, electrically programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), etc.), any suitable media that is not fleeting or devoid of any semblance of permanence during transmission, and/or any suitable tangible media. As another example, transitory computer readable media can include signals on networks, in wires, conductors, optical fibers, circuits, any suitable media that is fleeting and devoid of any semblance of permanence during transmission, and/or any suitable intangible media.
Referring back to FIG. 1, communication links 112 can be any suitable mechanism for connecting collaborating systems 102, 104, 106, masquerader/compromised computer 110, deception system 114, and attacking computer system 116 to communication network 108. Links 112 can be any suitable wired or wireless communication link, such as a T1 or T3 connection, a cable modem connection, a digital subscriber line connection, a Wi-Fi or 802.11(a), (b), (g), or (n) connection, a dial-up connection, and/or any other suitable communication link. Alternatively, communication links 112 can be omitted from system 100 when appropriate, in which case systems 102, 104, and/or 106, computer 110, and/or deception system 114 can be connected directly to communication network 108.
In some embodiments, detection and deception system 114 can be any computer, server, router, or other suitable device for generating models of normal user behavior and intent, monitoring user behavior in system 110, and modeling, generating, inserting, distributing, and/or managing decoy information into system 100. Similar to collaborating systems 102, 104, and 106, detection and deception system 114 can run programs, such as operating systems (OS), software applications, a library of functions and/or procedures, background daemon processes, and/or various other suitable programs. In some embodiments, detection and deception system 114 can support one or more virtual machines.
For example, detection and deception system 114 can include a decoy information broadcaster to inject decoy traffic information into communications network 108. In another example, detection and deception system 114 can be a designated server or a dedicated workstation that analyzes the information, events, and network flow in system 100, generates models of user behavior and decoy information based on that analysis, and inserts the deception information into the system 100. In yet another example, deception system can operate in connection with a host-based security application, such as Symantec Antivirus. In yet another example, detection and deception system 114 can be multiple servers or workstations that simulate the information, events, and traffic between collaborating systems 102, 104, and 106.
In some embodiments, detection and deception system 114 can include multiple sensors to monitor, audit, and/or capture data on the users of system 100. For example, detection and deception system 114 can include one or more sensors that monitor registry-based activities, process creation and destruction, window graphical user interface (GUI) access, dynamic link library (DLL) activity, and/or decoy information access (see, e.g., FIG. 2). Using such a sensor can obtain information on user activities, such as the process name and process identification, the process path, the parent of the process, the type of process action (e.g., the type of registry access, process creation, process destruction, etc.), the process command arguments, action flags (e.g., success or failure), and registry activity results. Each audit record or piece of information obtained by the sensor can be recorded with a timestamp. In a more particular example, the sensor can gather particular information based on the operating system used by the user. In a digital processing device using a Linux-based operating system, the sensor can use a kernel hook (e.g., the auditd daemon) to obtain process creation events and enriched user commands. In a digital processing device using a Microsoft Windows operating system, the sensor can use a low-level system driver, DLL registration mechanisms, and a system table hook to obtain registry-based activities, process creation and destruction, window GUI access, and DLL libraries activity. Examples of the different sensors are shown in FIG. 2. An illustrative example of a registry access sensor is shown, for example, in FIG. 3. As shown in screen 300, the sensor monitors and records process identification numbers, timestamps, scores, process names and command arguments, requests, process paths, results, and any other suitable information for a particular user.
As described above, masquerade detection mechanisms that profile user actions to generate models of user behavior and intent and that monitor user actions and access of decoy information can be provided to users of collaborating systems 102, 104, and/or 106, masquerader/compromised computer 110, and/or communications network 108 of FIG. 1. FIG. 4 illustrates an example 400 of a process for providing masquerade detection mechanisms in accordance with some embodiments of the disclosed subject matter. As shown, access is granted to a file system at 410. For example, access can be granted in response to receiving a legitimate username and password.
It should be noted that, in response to granting access to a file system, the monitored user behavior (at 420), the decoy information (at 430), selected categories, generated models of user behavior and intent (at 440), generated alerts (at 450), or any other suitable action described herein can be associated with a particular user or a particular user type. For example, the monitored user behavior (e.g., user commands, user-initiated applications, registry-based activities, and/or DLL activities) and models can be stored and associated with a particular user based on username. In another example, the models can be stored and associated with a particular user type, such as an administrator user or a guest user.
In response, user behavior on the file system is monitored at 420. Similarly, one or more sensors (e.g., Windows sensors, Linux sensors, or sensors for any suitable operating system) can be used to monitor information on user activities, such as the process name and process identification, the process path, the parent of the process, the type of process action (e.g., the type of registry access, process creation, process destruction, etc.), the process command arguments, action flags (e.g., success or failure), and registry activity results.
At 430, files that contain decoy information on the file system can be generated and monitored. It should be noted that one downside of a false positive is annoyance of a legitimate user who can be alerted too frequently. On the other hand, one downside of a false negative (e.g., an undetected masquerader) can be far more dangerous. These files containing decoy information (e.g., a decoy password file, a file with bogus credit card information, etc.) can be deployed and monitored. Along with the generation of a model of user behavior and intent and the comparison of current user actions (e.g., user commands, user-initiated applications, registry-based activities, and/or DLL activities) with the generated model, the detection of the performance of an action on a file containing decoy information provides stronger evidence of malfeasance, thereby improving the accuracy of detection and deception system 114. Accordingly, decoy information can be used to reduce and/or eliminate false positives.
Decoy information and/or decoy documents can be generated and/or managed using any suitable approach. For example, detection and deception system 114 can search through files on a computer (e.g., one or more of collaborating systems 102, 104, and 106), receive templates, files, or any other suitable input from a legitimate user (e.g., an administrator user) of a computer, monitor traffic on communications network 108, or use any other suitable approach to create believable decoy information. In a more particular example, detection and deception system 114 can determine which files are generally accessed by a particular user (e.g., top ten, last twenty, etc.) and generate decoy information similar to those files. In another example, detection and deception system 114 can perform a search and determine various usernames, passwords, credit card information, and/or any other sensitive information that may be stored on one or more of collaborating system 102, 104, and 106. Using these search results, detection and deception system 114 can then create receipts, tax documents, and other form-based documents with decoy credentials, realistic names, addresses, and logins. Alternatively, detection and deception system 114 can monitor the file system and generate decoy documents with file names similar to the files accessed on the file system (e.g., a tax document with the file name “2009 Tax Form-1099-1”) or with file types similar to the files accessed on the file system (e.g., PDF file, DOC file, URL link, HTML file, JPG file, etc.).
It should be noted that detection and deception system 114 can generate decoy information and decoy documents that comply with particular properties that enhance the deception for masqueraders. Decoy information can be generated such that the documents are believable, enticing, conspicuous, detectable, variable, differentiable from actual or authentic information, non-interfering with legitimate users, etc.
In some embodiments, decoy information can be combined with any suitable number of monitoring or alerting approaches, either internal or external, to detect masqueraders. For example, a beacon can be embedded in a document or any other suitable decoy information. As used herein, a beacon can be any suitable code or data that assist in the differentiation of decoy information from actual information and/or assists in indicating the malfeasance of an attacker illicitly accessing the decoy information. In some embodiments, these stealthy beacons can cause a signal to be transmitted to a server indicating when and/or where the particular decoy information was opened, executed, etc.
In another example, the decoy information, such as a decoy document, can be associated and/or embedded with one or more passive beacons. In a more particular example, a passive beacon in the form of a watermark can be embedded in the binary format of the document file or any other suitable location of the document file format. The watermark is detected when the decoy information is loaded in memory or transmitted in the open over a network (e.g., using the decoy file sensor of FIG. 2). In some embodiments, a host-based monitoring application can be configured to transmit signals or an alert when it detects the passive beacon in documents.
In yet another example, the content of the decoy information itself can be used to detect the potential presence of a masquerader. The content of the decoy information can include a bogus login (e.g., a bogus login and password for Google Mail). The bogus login to a website can be created in a decoy document and monitored by external approaches (e.g., polling a website or using a custom script that accesses mail.google.com and parses the bait account pages to gather account activity information).
Approaches for baiting inside attackers using decoy information are discussed in further detail in, for example, Keromytis et al., International Publication No. WO 2009/032379, published on Mar. 12, 2009, and Stolfo et al., U.S. Patent Application Publication No. 2010/0077483, (Attorney Docket No. 0315120.167-US3), published on Mar. 25, 2010, which are hereby incorporated by reference herein in their entireties.
Referring back to FIG. 4, user behavior and decoy information are monitored on the file system at 420 and 430. An illustrative process for generating models of user behavior and intent and determining that the current user actions are abnormal are described in further detail below in connection with FIG. 5. After determining that one or more of the current user actions are abnormal (e.g., determining that one or more of the current user actions deviate from the model of normal user behavior), it is also determined whether at least one of the current user actions involves accessing, executing, printing, copying, transmitting, or performing any other suitable action on a file that contains decoy information (a decoy document) at 440.
At 450, in response to determining that one or more of the current user actions are abnormal and in response to determining that at least one of the current user actions involves accessing, executing, printing, copying, transmitting, and/or performing any other suitable action on a file that contains decoy information, an alert is generated. For example, an alert component can transmit an email to the legitimate user of the file system. In response to receiving the alert (e.g., an email message), the legitimate user can be prompted to indicate whether the alert is a false alarm. In another example, an alert component can transmit a notification to an administrator user (e.g., an information technology professional). In yet another example, an alert component can transmit a message to a host sensor or a host-based monitoring application, such as an antivirus software application.
In accordance with some embodiments of the disclosed subject matter, models of user behavior and intent are generated and implemented using a process 500 as illustrated in FIG. 5. While masqueraders may be able to capture user credentials and/or confidential user information, these masqueraders are generally not able to capture or impersonate the behavior of the particular user. By generating models of user behavior and intent (e.g., using the normal pattern of submitted user actions) for a certain user, masqueraders can be accurately detected as their behavior is different from the behavior of the legitimate user. In some embodiments, the incorporation of decoy documents and monitoring actions performed on these decoy documents can reduce or eliminate false positives created by the one or more models.
It should be noted that, in some embodiments, user activities (e.g., user commands, user-initiated applications, registry-based activities, and/or DLL activities) can be placed into contextual categories. For example, user commands can be classified into one of multiple categories as certain categories of user commands can reveal user intent more than other categories. In another example, Windows applications can be classified into one of multiple categories of applications. In a more particular example, user search behavior can be an interesting group of user commands and/or command categories to monitor and profile since it indicates the user lacks information they are seeking. Again, a masquerader is unlikely to have the depth of knowledge of the victim's machine (e.g., files, locations of important directories, available applications, etc.). Accordingly, a masquerader is likely to first perform information gathering and search commands before initiating specific actions. In another example, remote access to other systems and the communication or egress of large amounts of data to remote systems can be an interesting group of user commands and/or command categories to monitor and profile since such commands and/or categories can provide an indication of illegal copying or distribution of sensitive information.
In yet another example, a legitimate user tends to initiate the same repeated actions or activities each time the user logs in to their environment before using it by initiating multiple applications (e.g., read e-mail, open a web browser, start a chat session, etc.). Similarly, the legitimate user tends to clean up and shut down particular applications and perform particular actions when the user logs off. Such repeated behaviors constitute a profile that can be modeled and used to check the authenticity of a user session before significant damage is done. Accordingly, certain categories from the taxonomy can reveal user intent for a particular user.
Referring back to FIG. 5, at 510, each action (e.g., user commands, user-initiated applications, registry-based activities, and/or DLL activities) of multiple actions can be assigned a category type. Category types can classify these actions or activities into particular categories, such as, for example, access control, applications, communications and networking, display and formatting, execution and program control, file system, I/O peripherals, search and information gathering, other, process management, system management, unknown, and utilities. Other examples of various categories are shown, for example, in the illustrative taxonomies of FIGS. 6 and 7.
As shown in FIGS. 6 and 7, some categories can be further divided or classified into sub-categories. For example, the display and formatting category can be further classified into file conversion commands or applications and editing commands or applications. In another example, the execution and program control category can be further classified into job execution commands and program control commands.
It should be noted that any other suitable category can also be created and that categories can be created based on, for example, operating environment (e.g., Windows, Unix, Chrome, etc.), user type (e.g., administrator, guest, etc.), etc. For example, FIG. 6 shows multiple categories created for a Unix command sequences issued by users, while FIG. 7 shows multiple categories created for a Windows operating environment with applications executed by users. In Unix systems, the search and information gathering category type can include user commands, such as find and fgrep, while the process management category type can include user commands, such as kill, nohup, and renice. In addition, the date, clock, and cal commands can be placed into the utilities category. The other and unknown categories can be used for commands that have been recognized but not classified under any other category and for commands that have not been identified or recognized, respectively.
Referring back to FIG. 5, a taxonomy of categories can be generated based on the assigned category type at 520. The generated taxonomy of categories, user commands, user-initiated applications, registry-based activities, and/or DLL activities can be used to readily identify and model particular user behavior. Examples of taxonomies for user commands in a Unix environment and for applications in a Windows environment are shown in FIGS. 6 and 7, respectively.
It should be noted that the taxonomy abstracts the audit data and enriches the meaning of a user's profile. That is, activities, such as commands, applications, or DLL activities, that perform similar types of actions are grouped together in one or more categories making profiled sequences more abstract and meaningful. A taxonomy can reduce complications, such as when “Never-Before-Seen-Commands” are observed, since distinct user commands are replaced by its category. Accordingly, user actions (e.g., user commands, user-initiated applications, registry-based activities, and/or DLL activities) are assigned a category type and the sequence of category types is modeled rather than individual actions.
At 530, a model of the user behavior and intent can be generated using, for example, audit data, monitored user actions, etc. As described above, some categories can reveal user intent more than other categories. Accordingly, particular categories that reveal user intent can be selected for modeling. For example, to reveal the intent of a particular user, categories relating to searching, access control privileges, and copying or printing information can be selected. Upon collecting user behavior from these categories (e.g., in a given window), a model can be generated of the normal pattern of submitted actions for one or more categories of the taxonomy for a particular user.
It should be noted, however, that user behavior varies for each user such that different categories can be selected for different users. For example, an administrative user can have a different set of available commands than a guest user. In another suitable embodiment, the detection and deception system 114 can prompt the user (e.g., the user that logged in, an administrative user, etc.) to input responses to questions relating to user behavior. For example, the user can be provided with an interface for inputting responses to questions relating to user search behavior (e.g., most frequent queries, particular folder locations, etc.). In another suitable example, an administrator user can be provided with an interface for selecting categories for particular users in particular divisions of a corporation. Alternatively, the detection and deception system 114 can select particular categories from the taxonomy based on user behavior (e.g., most frequent searches, top ten remote sources receiving transmissions of sensitive information, etc.).
Once these categories are identified, features representing such behavior can be extracted. For example, monitored data can be grouped into 10 second quanta of user activities, where seven features are extracted for each of these epochs—e.g., number of search actions (specific sections of the Windows registry, specific DLLs, and specific programs on the system are correlated with system search), number of non-search actions, number of user-induced actions (e.g., manually starting or killing a process, opening a window, manually searching for some file or some content, etc.), number of window touches (e.g., bringing a window into the foreground, closing a window, etc.), number of new processes, total number of processes running on the system, and total number of document editing applications running on the system. However, any suitable features can be extracted for each category.
Current user activities (e.g., real-time user actions) on the file system are monitored and compared to the generated model of “normal” user behavior and categories in order to determine its likelihood of being harmful at 540. Based on the comparison, it can be determined whether one or more of the monitored user actions are abnormal, thereby detecting a masquerader attack.
In some embodiments, generating a model of the submitted user behavior and comparing current user behavior with the model to determine its likelihood of being a masquerader attack in 530, 540, and 550 can be performed using Hellinger distance-based modeling approaches.
The Hellinger distance can be used to calculate the change in two frequency tables, where each table is a histogram representing the frequency of some variable at some particular moment in time. More particularly, in these embodiments, the frequency of user actions can be measured. Accordingly, a detector of abnormal behavior is created by modeling user action frequencies and the changes in that frequency.
The Hellinger distance can be defined as:
HD(f _p [ ],f _t[ ])=Σ_i=0 ^n-1(√{square root over (f _p [i])}−√{square root over (f _t [i])})²
where f_p[ ] is the array of normalized frequencies for the first set, f_t[ ] is the array of normalized frequencies for the second set, and n is the number of possible commands or categories. For example, the Hellinger distance metric can be applied in response to the issuance of an action (e.g., user commands, applications, registry-based activities, and/or DLL activities) by a particular user. A previous frequency table that modeled the previous actions in one or more categories is compared to a newly updated frequency table by modifying the frequency of the command types. Hence, each action creates a new or updated Hellinger distance score that is subjected to threshold logic. Each bin of the frequency table can be any category selected to model. In a general case, every category can be modeled and monitored.
In a more particular example, in order to detect changes in user behavior, the Hellinger distance between two blocks of actions of size w can be calculated. With each new action, the frequencies of the categories within the block of the last w commands can be counted. The window of w actions is then shifted by one action and the same frequencies for the new block can be counted.
In some embodiments, an average score over a particular number of blocks can be calculated to determine an overall score. In response to the average score being greater than a particular user threshold value, the block of actions (e.g., commands) can be deemed abnormal and a masquerader block.
Alternatively, in some embodiments, the rate of change of the Hellinger distance can be calculated. For example, a calculated Hellinger distance can be compared to a previously calculated Hellinger distance, where the difference between the Hellinger distances is determined. If the rate of change between the Hellinger distances is greater than a particular threshold value, the block of actions can be deemed abnormal and a masquerader block.
In some embodiments, generating a model of the submitted user behavior and comparing current user behavior with the model to determine its likelihood of being a masquerader attack in 530, 540, and 550 can be performed using support vector machine (SVM) modeling approaches.
Support vector machines (SVMs) are linear classifiers used for classification and regression and are known as maximal margin classifiers. Generally speaking, SVMs map input data into a high-dimensional feature space using a kernel function. The kernel function can be defined as:
k(x,y)=(Φ(x),Φ(y)),
where x, y∈X
X is the training data set and Φ is the feature mapping to a high-dimensional space X→F.
It should be noted that the feature mapping can be performed in any suitable manner. For example, in some embodiments, a frequency-based approach can be conducted, where the number of times a category retrieved using the taxonomy appears in the monitored data is counted. In another example, in some embodiments, a binary approach can be conducted, where the presence of the category in the monitored data is indicated.
Accordingly, methods, systems, and media are provided for detecting masquerade attacks by monitoring computer user behavior.
Although the invention has been described and illustrated in the foregoing illustrative embodiments, it is understood that the present disclosure has been made only by way of example, and that numerous changes in the details of implementation of the invention can be made without departing from the spirit and scope of the invention. Features of the disclosed embodiments can be combined and rearranged in various ways.

Claims

What is claimed is:

1. A method for detecting attacks on devices, the method comprising:

receiving, using a hardware processor, a first plurality of user actions in a computing environment associated with a user account, wherein the first plurality of user actions includes a user selection of one or more particular documents in the computing environment;

generating a model of user behavior based at least in part on the first plurality of user actions, wherein the model of user behavior is associated with the user account;

receiving a second plurality of user actions in the computing environment that correspond to the user account;

determining whether at least one of the second plurality of user actions deviates from the generated model of user behavior, wherein the determination further comprises determining whether the second plurality of user actions includes accessing the one or more particular documents in the computing environment;

identifying whether the second plurality of user actions is an attack based at least in part on the determination that at least one of the second plurality of user actions deviates from the generated model of user behavior and that the one or more particular documents in the computing environment has been accessed; and

generating an alert of the attack in response to the identification.

2. The method of claim 1, further comprising searching through the computing environment to identify the one or more particular documents.

3. The method of claim 2, further comprising identifying the one or more particular documents by determining that the one or more particular documents from a plurality of documents are commonly accessed by a user corresponding to the user account.

4. The method of claim 2, further comprising identifying the one or more particular documents by determining that the one or more particular documents from a plurality of documents are commonly accessed by a plurality of user accounts associated with the computing environment.

5. The method of claim 1, further comprising:

generating decoy documents that are similar to the one or more particular documents; and

inserting the decoy documents into the computing environment, wherein identifying that the second plurality of user actions is the attack includes detecting that at least one of the generated decoy documents has been accessed.

6. The method of claim 1, further comprising:

generating documents that are similar to the one or more particular documents;

inserting decoy information into each of the generated documents; and

inserting the generated documents into the computing environment, wherein identifying that the second plurality of user actions is the attack includes detecting that at least one of the generated documents has been accessed.

7. The method of claim 6, wherein the decoy information further comprises a beacon that is configured to indicate when a document associated with the beacon has been accessed.

8. The method of claim 6, further comprising transmitting a monitoring instruction to a monitoring application that monitors a beacon in the decoy information to detect when a document associated with the beacon has been accessed.

9. The method of claim 1, further comprising

classifying each of the first plurality of user actions into a category of a plurality of categories of user actions; and

generating the model of user behavior based on user actions in at least one category of the plurality of categories of user actions.

10. The method of claim 9, further comprising receiving a selection of at least one category of the plurality of categories of user actions for generating the model of user behavior.

11. The method of claim 1, further comprising assigning a category type to each of a plurality of user commands, applications, registry-based activities, and dynamic link library (DLL) activities in the computing environment.

12. The method of claim 11, further comprising:

generating a taxonomy of categories based on the category type;

selecting one or more categories from the taxonomy;

extracting a plurality of features for each category; and

generating the model of user behavior by using the first plurality of user actions with respect to the extracted features.

13. The method of claim 1, further comprising calculating a first distance between a first distribution that models the first plurality of user actions and a second distribution that models the second plurality of user actions.

14. The method of claim 13, further comprising determining that at least one of the second plurality of user actions deviates from the model of user behavior in response to the first distance being greater than a threshold distance value.

15. The method of claim 13, further comprising:

calculating a second distance between the second distribution that models the second plurality of user actions and a third distribution that models a third plurality of user actions; and

comparing the second distance with the first distance to determine whether the third plurality of user actions corresponds to the attack.

16. The method of claim 1, further comprising associating the model of user behavior with at least one of: a particular user, the computing environment, a network, and a user type.

17. A method for detecting attacks on devices, the method comprising:

receiving, using a hardware processor, a first plurality of user actions in a computing environment associated with a user account;

generating a model of user behavior based on the first plurality of user actions, wherein the model of user behavior is associated with the user account;

receiving additional user actions in the computing environment that correspond to the user account;

determining whether at least one of the additional user actions deviates from the generated model of user behavior; and

in response to determining that at least one of the additional user actions deviates from the generated model of user behavior, generating an alert that includes information associated with the user account.

18. The method of claim 17, further comprising determining that the additional user actions include performing search activities in the computing environment, wherein the alert is generated in response to determining that the additional user actions include the search activities in the computing environment.

19. The method of claim 18, wherein the search activities include a number of search actions conducted in the computing environment in a given period of time.

20. The method of claim 17, further comprising:

21. The method of claim 17, further comprising assigning a category type to each of a plurality of user commands, applications, registry-based activities, and dynamic link library (DLL) activities in the computing environment.

22. The method of claim 21, further comprising receiving a selection of at least one category of the plurality of categories of user actions for generating the model of user behavior.

23. The method of claim 21, further comprising:

generating a taxonomy of categories based on the category type;

selecting one or more categories from the taxonomy;

extracting a plurality of features for each category; and

24. The method of claim 21, further comprising calculating a first distance between a first distribution that models the first plurality of user actions and a second distribution that models the second plurality of user actions.

25. The method of claim 24, further comprising determining that at least one of the second plurality of user actions deviates from the model of user search behavior in response to the first distance being greater than a threshold distance value.

26. The method of claim 24, further comprising:

27. The method of claim 17, further comprising associating the model of user search behavior with at least one of: a particular user, the computing environment, a network, and a user type.

28. A method for detecting attacks on devices, the method comprising:

monitoring, using a hardware processor, a first plurality of user actions over a communications network;

generating a model of user behavior based on the first plurality of user actions over the communications network;

receiving additional user actions over the communications network;

in response to determining that at least one of the additional user actions deviates from the generated model of user behavior, generating an alert that includes information associated with the communications network.

29. The method of claim 28, wherein the first plurality of user actions and the second plurality of user actions includes networking-related user activities.

30. The method of claim 28, wherein the first plurality of user actions and the second plurality of user actions includes communications-related user activities.

31. The method of claim 28, further comprising:

classifying each of the first plurality of user actions into a category of a plurality of categories of user actions, wherein the plurality of categories of user actions includes a first category of communications-related user activities and networking-related user activities;

extracting user actions from the first plurality of user actions that have been placed in the first category; and

generating the model of user behavior based on the user actions in the first category of the plurality of categories of user actions.

32. The method of claim 28, further comprising associating the model of user behavior with the communications network.