Detecting malware based on reflection

Embodiments of the disclosure relate to the field of cyber security. More specifically, embodiments of the disclosure relate to a system and method for detecting malware utilizing reflection for obfuscation.

Malicious software (“generally referred to as “malware”) has become a pervasive problem for corporations and individual users alike, as the functionality of most networked resources is based on downloaded software. The presence of malware within downloaded software may compromise a networked resource and the network itself. A number of techniques have been used by malware authors to obfuscate the analysis of their malware within downloaded content.

Currently, security appliances are not equipped to consistently detect malware when obfuscated by malware authors using advanced programmatic techniques.

Embodiments of the invention are illustrated by way of example and not by way of limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:

FIG. 1 is a first illustrative embodiment of a threat detection platform (TDP) deployed within a network that detects malware that uses reflection for obfuscation.

FIG. 2 is a second illustrative embodiment of the TD) deployed within a network that operates in combination with remote sources to detect malware that uses reflection for obfuscation.

FIG. 3 is an exemplary embodiment of a logical representation of the TDP of FIG. 1 .

FIG. 4 is a general exemplary flowchart that illustrates operations conducted by one or more electronic devices for determining whether an object that invokes reflection operations is malicious.

FIG. 5 is a first exemplary flowchart that illustrates operations collectively conducted by a static analysis engine and a classification system for determining whether an object invoking reflection operations is malicious.

FIG. 6 is a second exemplary flowchart that illustrates operations collectively conducted by a dynamic analysis engine and a classification engine for determining whether an object that invokes reflection operations is malicious.

FIG. 7 is an exemplary flowchart of the operations of the classification analysis performed by the classification engine of FIGS. 1 and 2 .

Various embodiments of the disclosure relate to a platform that is implemented with logic configured to (i) analyze the content of an object to determine whether the object is configured to issue a function call that invokes reflection operations, and/or (ii) detect whether the object, when processed, issues a function call that invokes reflection operations. The functionality of this logic is directed to uncover malware that relies on reflection for obfuscation purposes.

In general, “reflection” represents an ability to examine or modify run-time behaviors of a particular object. As an example, in object oriented programming languages such as JAVA®, reflection allows for inspection of software components, such as interfaces as well as source code constructions (e.g., classes) at run-time, without knowing the names of these software components at compile time.

As an illustrative embodiment, such detection may involve a determination as to whether an object under analysis (sometimes referred to as a “suspect object”) is configured to or is attempting to access one or more application programming interfaces (APIs) that invoke reflection operations (hereinafter “reflection APIs”). In response to determining that the object is configured to or is attempting to access a reflection API, an analysis of one or more features of the object may be conducted to determine whether the object may be associated with a malicious attack. This analysis may involve probabilistic modeling analysis and/or machine learning analysis, as described below.

More specifically, a threat detection platform (TDP) may be deployed to conduct a first analysis of a suspect object to determine whether the suspect object is configured to issue a function call that invokes reflection operations, such as an API call to a reflection API for example. According to one embodiment of the disclosure, a static analysis engine of the TDP may be configured to conduct an operation (e.g., de-obfuscation such as decompiling and/or disassembling incoming data or even emulation) to recover content from the suspect object. The content may be part of a high-level representation of the object, such as at least a portion of source code, pseudo-code, or another human readable format. Thereafter, the content may be analyzed in efforts to detect the presence of one or more function calls that, during run-time, would invoke reflection operations.

For example, the static analysis engine may be configured to decompile an object, such as an executable file for example, to recover source code. Thereafter, the static analysis engine analyzes the source code to determine if the source code includes a function call that invokes reflection operations. For instance, the source code may include an API call to a predetermined reflection API. Upon completion of a scan of the source code (e.g., an examination without execution) and detection of a function call that invokes reflection operations (e.g., an API call to a reflection API), the object is determined to be suspicious. The object is deemed “suspicious” when there exists at least a first level of likelihood of the object being associated with a malicious attack.

Additionally, or in the alternative, reflection can be identified by implementing logic within a dynamic analysis engine of the TDP. During virtual processing of the suspect object, the logic may be adapted to set interception points (e.g., hooks, breakpoints, etc.) that are used to detect the presence of one or more function calls that invoke reflection operations (e.g., particular API or system calls, etc.). Hence, in response to detecting a function call that invokes reflection operations, logic within the dynamic analysis engine determines that the object is “suspicious”.

After the object is deemed “suspicious” in response to determining that the content associated with the object includes a function call or determining that the object issues a function call that invokes reflection operations, the static analysis engine and/or the dynamic analysis engine provides the suspicious object and/or particular features associated with the suspicious object to the classification system for a more in-depth analysis. Deployed within the security appliance or in a remotely located resource, the classification system is configured to determine whether the suspicious object is “malicious,” namely the system determines whether there is a prescribed likelihood (higher than the first level of likelihood) of the object being associated with a malicious attack. In general, it is contemplated that the classification system may not be accessed unless the suspect object (i) is configured to issue a function call that invokes reflection operations or (ii) has issued a function call that invokes reflection operations.

According to one embodiment of the disclosure, the classification system determines whether the object is malicious by applying a probabilistic model analysis to one or more features (herein “feature(s)”) extracted from the suspicious object after analysis by the static analysis engine and/or the dynamic analysis engine. These feature(s) may include, but are not limited or restricted to metadata (e.g., function names and/or object size), parameters passed (or to be passed) with an intended function call, and/or other information potentially indicative of malware such as suspicious data strings from content of the object if the object has been successfully de-obfuscated. It is contemplated that the feature(s) may further include information associated with behaviors that constitute abnormalities such as a reflection API downloading a file or executing a file.

The classification system may, in addition, or in the alternative, apply a machine learning analysis to the feature(s) associated with the suspicious object. Machine learning analysis includes an operation of comparing the feature(s), either individually or as a pattern of two or more features, to data that is known to be malicious or non-malicious (e.g., benign). This comparison determines whether the suspicious object is malicious or non-malicious.

The results of the probabilistic model analysis, the machine learning analysis, or a combination of these analyses produces a result that identifies whether the suspicious object is deemed to be malicious or non-malicious.

I. Terminology

In the following description, certain terminology is used to describe aspects of the invention. For example, in certain situations, both terms “logic” and “engine” are representative of hardware, firmware and/or software that is configured to perform one or more functions. As hardware, logic (or engine) may include circuitry having data processing or storage functionality. Examples of such processing or storage circuitry may include, but is not limited or restricted to a (hardware) processor; one or more processor cores; a programmable gate array; a microcontroller; an application specific integrated circuit; receiver, transmitter and/or transceiver circuitry; storage medium including semiconductor memory or a drive; or combinatorial logic, or combinations of one or more of the above components.

Logic (or engine) may be in the form of one or more software modules, such as executable code in the form of an executable application, an application programming interface (API), a subroutine, a function, a procedure, an applet, a servlet, a routine, source code, object code, a shared library or dynamic-link library (dll), or one or more instructions. These software modules may be stored in any type of a suitable non-transitory storage medium, or transitory storage medium (e.g., electrical, optical, acoustical or other form of propagated signals such as carrier waves, infrared signals, or digital signals). Examples of a “non-transitory storage medium” may include, but are not limited or restricted to a programmable circuit; non-persistent storage such as volatile memory (e.g., any type of random access memory “RAM”); persistent storage such as non-volatile memory (e.g., read-only memory “ROM”, power-backed RAM, flash memory, phase-change memory, etc.), a solid-state drive, hard disk drive, an optical disc drive, or a portable memory device; and/or a semiconductor memory. As firmware, the executable code is stored in persistent storage.

The term “object” generally refers to a collection of data, whether in transit (e.g., over a network) or at rest (e.g., stored), often having a logical structure or organization that enables it to be classified for purposes of analysis. For instance, the object may be a file (e.g., Portable Document Format “PDF” document, or Microsoft® Word® or other word processing document), or HyperText Markup Language “HTML” based web page, or the like. During analysis, for example, the object may exhibit or a program processing the object may exhibit one or more behaviors that are systematic of malicious activity and provide evidence that the object may be classified as malicious. One of these behaviors may include issuance of a function call that invokes one or more reflection operations.

One example of a function call that invokes reflection operations is an API call to access a reflection API (e.g., an API call to “Class.forName(X)” that causes the class named “X”, namely a programming construct with particular function to be dynamically loaded at run-time). Another example of a function call that invokes reflection operations may be a system call, normally based on an API call, where the called system function invokes reflection operations.

A “platform” generally refers to an electronic device with connectivity to an external data source (e.g., network, other electronic device, etc.) that typically includes a housing that protects, and sometimes encases, circuitry with data processing and/or data storage. Examples of a platform may include a server, a dedicated security appliance, or an endpoint device which may include, but is not limited or restricted to a stationary or portable computer including a desktop computer, laptop, netbook or tablet; a smartphone; a video-game console; or wearable technology (e.g., smart watch, etc.).

The term “transmission medium” is a physical or logical communication path with an endpoint device. For instance, the communication path may include wired and/or wireless segments. Examples of wired and/or wireless segments include electrical wiring, optical fiber, cable, bus trace, or a wireless channel using infrared, radio frequency (RF), or any other wired/wireless signaling mechanism.

The term “computerized” generally represents that any corresponding operations are conducted by hardware in combination with software and/or firmware.

Lastly, the terms “or” and “and/or” as used herein are to be interpreted as inclusive or meaning any one or any combination. Therefore, “A, B or C” or “A, B and/or C” mean “any of the following: A; B; C; A and B; A and C; B and C; A, B and C.” An exception to this definition will occur only when a combination of elements, functions, or operations are in some way inherently mutually exclusive.

As this invention is susceptible to embodiments of many different forms, it is intended that the present disclosure is to be considered as an example of the principles of the invention and is not intended to limit the invention to the specific embodiments shown and described.

II. General Architecture

Referring to FIG. 1 , an exemplary block diagram of a network 100 deploying a plurality of threat detection platforms (TDP) 110 1 - 110 N (N>1, where N=3 for this embodiment) communicatively coupled to a management system 120 via a network 125 is shown. In general, the management system 120 is adapted to manage each TDP 110 1 - 110 3 . For instance, the management system 120 may be configured to perform content updates within a processing engine operating as part of a communication interface 135 , a static analysis engine 140 , a dynamic analysis engine 160 , a classification engine 180 , and/or a reporting engine 190 with an optional user interface capability. For example, the content update may include a software or firmware update that alters the functionality of the TDP 110 1 . Alternatively, the content update may include security content, such as signatures or rules changes (e.g., add/delete/modify signatures, rules or parameters that are utilized by the rules, etc.). The static analysis engine 140 and/or the dynamic analysis engine 160 may use the signatures and/or rules to detect whether reflection operations are invoked and whether the reflection operations are directed to malicious activities.

As shown in FIG. 1 , a first threat detection platform (TDP) 110 1 is an electronic device that is adapted to analyze information associated with incoming data (e.g., network traffic propagating over a communication network 132 , input data from another type of transmission medium including a dedicated transmission medium, etc.). As this illustrative embodiment, the first TDP 110 1 is communicatively coupled with the communication network 132 via an interface 136 , where the communication network 132 may include a public network such as the Internet, a private network (e.g., a local area network “LAN”, wireless LAN, etc.), or a combination thereof. The interface 136 operates as a data capturing device that intercepts (or alternatively duplicates) at least a portion of the data associated with an object, which may include metadata. Alternatively, although not shown, the interface 136 may be configured to receive files or other objects that are not provided over a network. For instance, as an example, the interface 136 may be a data capturing device that automatically (or on command), accessing data stored in a storage system or another type of interface, such as a port, for receiving objects manually provided via a suitable dedicated communication link or from storage media such as portable flash drives.

In some embodiments, although not shown, interface 136 may be contained within the first TDP 110 1 . In other embodiments, the interface 136 can be integrated into an intermediary device in the communication path (e.g., an optional firewall 137 , router, switch or other networked electronic device) or can be a standalone component, such as an appropriate commercially available network tap.

For this illustrative embodiment, however, the interface 136 may be configured to capture data associated with an incoming object for analysis, and perhaps its corresponding metadata (or generate metadata based on the captured data). The metadata may be used, at least in part, to determine protocols, application types and other information that may be used by logic (e.g., scheduler 150 or a virtual machine monitor not shown) within the first TDP 110 1 to determine particular software profile(s) used for virtual machine (VM) configuration and/or VM operation scheduling. For instance, the software profile(s) may be used for selecting and/or configuring one or more virtual machines (VMs) 163 1 - 163 M (M≧1) within a virtual analysis environment 162 of the dynamic analysis engine 160 . These software profile(s) may be directed to different software or different versions of the same software application extracted from software image(s) fetched from a storage device 155 . Additionally, the metadata may be used, at least in part, as the feature(s) that are evaluated by a classification system 182 within the classification engine 180 in determining whether the object under analysis is malicious or not.

As further shown in FIG. 1 , the first TDP 110 1 includes communication interface 135 , static analysis engine 140 , scheduler 150 , storage device 155 , dynamic analysis engine 160 , classification engine 180 , and/or reporting engine 190 . Herein, according to this embodiment of the disclosure, the communication interface 135 receives an object and converts that object into a format, as need or appropriate, on which scanning may be conducted by the static analysis engine 140 (see operation ( 1 )). This conversion may involve decompression of the object for example. It is contemplated that the communication interface 135 may conduct decompilation, disassembly or other de-obfuscation activities on the object and/or extraction of specific data associated with the object; however, according to this embodiment as described below, the de-obfuscation and data extraction activities may be handled by logic within the static analysis engine 140 .

As shown in FIG. 1 , the static analysis engine 140 comprises de-obfuscation logic 142 , reflection API analysis logic 144 , and/or feature extraction logic 146 (and their collective operations are illustrated as operation ( 2 )). The de-obfuscation logic 142 is configured to de-obfuscate at least a portion of an incoming object received from the communication interface 135 . As an example, the de-obfuscation logic 142 may be configured to de-obfuscate, such as decompile and/or disassemble, at least a portion of the incoming object (e.g., an executable) to recover a high-level representation of the object. The high-level representation may be in the form of source code, pseudo-code, or another high-level language.

After de-obfuscation, the reflection API analysis logic 144 may analyze content that is part of the high-level representation of the object for the presence of one or more API calls to any reflection API. In response to determining that the suspect object includes content that, at run-time, would issue an API call to one of the reflection APIs, the feature extraction logic 146 may extract feature(s) from the high-level representation (e.g., source code, or pseudo-code or another high-level language), such as called function names, data associated with the size of the object, information associated with one or more post infection behaviors, or the like. According to this embodiment of the disclosure, the extracted feature(s) may be provided as static analysis (SA)-based results 145 to the classification system 182 of the classification engine 180 for subsequent analysis.

It is contemplated that the static analysis engine 140 may further include processing circuitry that is responsible for extracting and/or generating metadata contained within or otherwise associated with incoming data from the communication interface 135 (e.g., network traffic, downloaded data). This metadata may be subsequently used for configuring one or more VMs 163 1 - 163 M within a virtual analysis environment 162 for conducting a dynamic analysis of the object 148 associated with that metadata.

Referring still to FIG. 1 , the reflection API analysis logic 144 of the static analysis engine 140 analyzes content within the object, which may be a portion of network traffic (or downloaded data) according to this embodiment of the disclosure. Such analysis may involve the performance of one or more checks on content associated with the object, namely content that is part of the high-level representation of the object, without execution of the object. Examples of the checks may include signature checks, which may involve a comparison of content that is part of the high-level representation of the object to one or more pre-stored signatures, which may include one or more reflection API function names.

After scanning the content of the suspect object, the reflection API analysis logic 144 determines whether or not this object is “suspicious” based on whether content within the high-level representation includes an API call to a reflection API. As a result, the static analysis engine 140 may pass this suspicious object 148 to the dynamic analysis engine 160 for more in-depth analysis in a VM-based analysis environment 162 (see operation ( 3 )). Additionally, or in the alternative, the reflection API analysis logic 144 may signal the feature extraction logic 146 to obtain one or more features associated with the suspect object and provide such feature(s) 143 to the classification engine 180 as part of SA-based results 145 (see operation ( 4 )).

Additionally, after analysis of the object has been completed, the static analysis engine 140 may provide some or all of the incoming object as the suspicious object 148 to the dynamic analysis engine 160 for in-depth dynamic analysis by one or more VMs 163 1 - 163 M of the virtual analysis environment 162 . For instance, according to one embodiment of the disclosure, a first VM 163 1 may be adapted to process the suspicious object 148 . Logic within the dynamic analysis engine 160 (e.g., reflection hooking logic 165 within the first VM 163 1 ) may be configured to monitor for certain types of behaviors exhibited by the suspicious object 148 during processing within the first VM 163 1 . One type of behavior may include the object 148 invoking reflection operations through one or more API calls to a reflection API. Another type of behavior may include detection of a system call (or, where a virtualization layer include a hypervisor is employed in an embodiment, a hyper call) that invokes reflection operations, where the system call (or hyper call) may be issued (or triggered) by the suspicious object 148 at run-time and may be based on an API call.

Herein, according to one embodiment, the first VM 163 1 is configured to process the suspicious object 148 . The reflection hooking logic 165 may be used to set one or more hooks at one or more reflection APIs or equivalent operating system (e.g., guest or host OS) functions that may perform or invoke reflection operations, where the hooks redirect the operational flow such as redirecting operations via a JUMP instruction to the classification system as described below (see operation ( 5 )). Examples of these reflection APIs may include, but are not limited or restricted to getClass( ) API or Class.forName( ), which are responsible for finding a class associated with the object.

Upon determining that the object 148 is issuing function calls to access an API or OS function that invokes reflection operations, the object feature extraction logic 167 may be activated to extract one or more features 172 (e.g., arguments, etc.) from the function call(s). Similarly, these feature(s) 172 may include a name of the function identified in the function call and/or other data within the arguments of the function call issued (or triggered) by the object 148 during processing within the first VM 163 1 . The feature(s) 172 may be stored in data store 170 and are subsequently provided to (or accessible by) the classification system 182 as part of VM-based results 175 .

Referring still to FIG. 1 , the scheduler 150 may be adapted to configure one or more VMs 163 1 - 163 M based on metadata associated with the suspicious object 148 in order to conduct run-time processing of the suspicious object 148 within the configured VMs 163 1 - 163 M . For instance, the first VM 163 1 and a second VM 163 2 may be configured to run concurrently (i.e. overlapping at least in part in time), where each VM 163 1 and 163 2 being configured with a different software profile corresponding to software images stored within storage device 155 . As an alternative embodiment, the first VM 163 1 may be configured to run plural processes concurrently or sequentially, each process configured according to a software configuration that may be used by different electronic devices connected to a particular enterprise network (e.g., endpoint device(s) 130 ) or a prevalent type of software configuration (e.g., a particular version of Windows® OS and/or a particular version of a web browser with a particular application plug-in). It is contemplated that the VM configuration described above may be handled by logic other than the scheduler 150 .

According to one embodiment of the disclosure, the dynamic analysis engine 160 may be adapted to execute one or more VMs 163 1 - 163 M that each simulate processing of the suspicious object 148 within a run-time environment. For instance, dynamic analysis engine 160 may include processing logic 161 to provide anticipated signaling to the VM(s) 163 1 , . . . , and/or 163 M during virtual processing of the suspicious object 148 , and as such, emulate a source of and/or destination for communications with the suspicious object 148 while processed within the VM(s) 163 1 , . . . , and/or 163 M . As an example, the processing logic 161 may be adapted to operate by providing simulated key inputs from a keyboard, keypad or touch screen, as requested by the suspicious object 148 during run-time.

Referring still to FIG. 1 , the static analysis engine 140 may be adapted to provide SA-based results 145 to the classification system 182 while the dynamic analysis engine 160 may be adapted to provide the VM-based results 175 to the classification system 182 (see operations ( 4 , 6 )). According to one embodiment of the disclosure, the SA-based results 145 may include information obtained by analyzing the incoming object that is potentially indicative of malware (e.g., function names, object size, suspicious strings within the object 148 ). Similarly, the VM-based results 175 may include information associated with the object 148 as well as the function calls that invoke reflection operations (e.g., function names or other argument data associated with the functions calls).

According to one embodiment of the disclosure, the classification engine 180 includes the classification system 182 that is configured to receive the SA-based results 145 and/or the VM-based result 175 associated with the object under analysis. Based at least partially on the SA-based results 145 and/or VM-based results 175 , the classification system 182 evaluates the feature(s) within the SA-based results 145 and/or VM-based results 175 to determine whether the suspicious object 148 should be classified as “malicious” (see operation ( 7 )).

For instance, as an illustrative embodiment, the SA-based results 145 include one or more features that are provided to probabilistic modeling logic 184 . The probabilistic modeling logic 184 is configured as a decision-tree analysis scheme, which receives one or more features as input, either individually or as a pattern of two or more features, and produces a result that may be used to identify whether the object is associated with a malicious attack.

According to one embodiment, the result may identify a risk level that indicates a likelihood of the object being associated with a malicious attack. For instance, the risk level may be identified in a variety of manners. For instance, the risk level may be conveyed by a two-state result that simply represents the object as malicious or non-malicious. Another risk level may be conveyed through a tri-state result (high, medium, low) to identify various probabilities of the object being associated with the malicious attack and obfuscated by reflection. Yet another risk level may be conveyed using scores that provide a greater granularity as to the likelihood of the object being associated with a malicious attack and obfuscated by reflection.

As an illustrative example, the result may include an overall score that is formed by an aggregation of scores (e.g., prescribed values) for some or all of the features undergoing analysis by the probabilistic modeling logic 184 . Herein, the name of a function call directed to a particular reflection API that is detected within the de-obfuscated content of the object may be assigned a first score. Similarly, the name of a system function that invokes reflection operations and is extracted from a system call detected during virtual processing of the object 148 may be assigned a second score different than the first score. Again, the size of the object may be assigned a third score, which is different than the first and second scores. The aggregation of these scores may be used to compute an overall score, which represents the likelihood of the object being malware that is obfuscated through reflection.

As an illustrative example, suppose that the object under static analysis is a file having a filename entitled “2014_IRS_TAX_INQUIRY” with a size of 15 megabytes and including content that represents a function call to a reflection API (e.g., getClass( )). According to this probabilistic modeling analysis, an aggregate value (e.g., a score greater than or equal to 8 out of a maximum 10) denotes that the object 148 is malicious. The probabilistic model logic 184 may include a portion of the decision-tree analysis that includes the following:

If object_content_string=string ‘getClass’

Score+=Score+4;

If filename string>8

if first_char=char ‘[0-9]’

Score+=Score+2.5

if first_char=char ‘[A-Z]”

Score+=Score+1.5

If filesize>10 megabytes

Score+=Score+2

Based at least in part on the one or more features associated with the object, a determination may be made by the probabilistic modeling logic 184 of the classification system 182 as to whether or not the object that invokes reflection is associated with a malicious attack. Upon determining that the object is associated with a malicious attack, the classification system 182 may provide information to identify the malicious object, including the resultant score and/or one or more of the features provided as part of the SA-based results 145 , to the reporting engine 190 .

As another illustrative embodiment, if provided in lieu of or in addition to SA-based results 145 , the VM-based results 175 may include one or more features 172 that are provided to probabilistic modeling logic 184 based on monitored behaviors during processing of the object 148 within the first VM 163 1 . According to this illustrative example, the probabilistic model logic 184 assigns a risk level to the object 148 under dynamic analysis. For a file having a filename (2014_IRS_TAX_INQUIRY) with a size of 15 megabytes and including content (e.g., a code that initiates a function call to access the reflection API such as getClass( )), the probabilistic modeling logic 184 may assign a risk level (e.g., aggregate score of at least 8 out of a maximum 10) that denotes that the object 148 is malicious. For this example, the probabilistic model logic 184 may include a portion of the decision-tree analysis that includes the following:

For this illustrated embodiment, based at least in part on the feature(s) associated with the object 148 , a determination may be made by the probabilistic modeling logic 184 of the classification engine 180 as to whether or not the object 148 is associated with a malicious attack. Upon determining that the object 148 is associated with a malicious attack (when Score≧8), the classification engine 180 may provide information to identify the malicious object, including one or more of the features 172 or the resultant score, to the reporting engine 190 .

As shown in FIG. 1 , the reporting engine 190 is configured to receive information 185 from the classification engine 180 and generate alerts 192 , especially in response to the suspicious object being now classified as malicious (see operation ( 8 )). The alerts may include various types of messages, which may include text messages and/or email messages, video or audio stream, or other types of information over a wired or wireless communication path. The reporting engine 190 features an optional user interface 194 (e.g., touch pad, keyed inputs, etc.) for customization as to the reporting configuration.

In addition, or in the alternative to probabilistic modeling logic 184 , the classification engine 180 may comprise machine learning logic 186 . Machine learning logic 186 performs an analysis of the one or more features that are part of the SA-based results 145 and/or the one or more features that are part of the VM-based results 175 . These features are compared, either individually or as a pattern of two or more features, to data known to be malicious or non-malicious (e.g. benign). The comparison is conducted to determine whether the object under analysis is malicious. Upon determining that the object is malicious (i.e., associated with a malicious attack), the classification engine 180 may provide information to identify the malicious object, such as one or more of the features from the SA-based results 145 and/or the VM-based result 175 and/or resultant score, to the reporting engine 190 .

Referring now to FIG. 2 , according to another embodiment of the disclosure, the static analysis engine 140 and/or dynamic analysis engine 160 located within the first TDP 110 1 may determine that the object is suspicious when the object is configured to invoke or invokes reflection operations, as described above (see operations 1 - 2 & 4 - 5 ). However, located remotely from the first TDP 110 1 , such as part of a cloud computing service 138 or within a different enterprise network for example, a classification system 200 is configured to receive an identifier 210 for the object along with (i) the object 137 and/or one or more features 143 ; (ii) object 148 and/or one or more features 172 ; or any combination thereof (see operations 3 , 6 and 7 ). The identifier 210 may include any value that is considered to be unique, such as a hash result (e.g., MD5 hash value) for example.

Including the probabilistic modeling logic 184 and/or machine learning logic 186 , the classification system 200 determines whether the object 148 is malicious and returns a result 220 of its probabilistic analysis or machine learning analysis (described above) along with the identifier 210 to the classification engine 180 (see operation 8 ).

Upon determining that the object 137 or 148 is associated with a malicious attack, the classification engine 180 may provide information 230 to identify the malicious object, including one or more of the features 143 or 172 and/or the result 220 (e.g., resultant score value), to the reporting engine 190 . Upon determining that the object 137 or 148 is benign, the classification engine 180 may provide information 230 to identify the object and that the object is benign, including the result 220 , to the reporting engine 190 . In lieu of reporting benign objects, the classification engine 180 may merely report malicious objects to the reporting engine 190 (see operation 9 ).

As still shown in FIG. 2 , the reporting engine 190 is configured to receive information from the classification engine 180 and generate alerts 192 , especially in response to the suspicious objects that have now been classified as malicious (see operation 10 ).

III. Exemplary Logic Layout of TDP

Referring now to FIG. 3 , an exemplary embodiment of a logical representation of the first TDP 110 1 is shown. The first TDP 110 1 includes a housing 305 , which is made entirely or partially of a rigid material (e.g., hardened plastic, metal, glass, composite or any combination thereof) that protect circuitry within the housing 305 , namely one or more processors 300 that are coupled to communication interface logic 310 that is part of communication interface 135 of FIGS. 1-2 via a first transmission medium 320 . Communication interface logic 310 enables communications with other TDP 110 2 - 110 3 and management system 120 of FIG. 1 . According to one embodiment of the disclosure, communication interface logic 310 may be implemented as a physical interface including one or more ports for wired connectors. Additionally, or in the alternative, communication interface logic 310 may be implemented with one or more radio units for supporting wireless communications with other electronic devices.

Processor(s) 300 is further coupled to persistent storage 340 via a second transmission medium 330 . According to one embodiment of the disclosure, persistent storage 340 may include (a) static analysis engine 140 , including de-obfuscation logic 142 , reflection API analysis logic 144 and feature extraction logic 146 ; (b) the dynamic analysis engine 160 that includes the processing logic 161 and the virtual analysis environment 162 that includes VM(s) 163 1 - 163 M , where at least some of the VM(s) 163 1 - 163 M include reflection hooking logic 165 and object feature extraction logic 167 ; (c) classification engine 180 ; (d) reporting engine 190 ; and/or (e) one or more data stores 350 that may be utilized by static analysis engine 140 , dynamic analysis engine 160 , classification engine 180 , and/or reporting engine 190 . One or more of these engines (or logic units) could be implemented externally from the first TDP 110 1 .

Collective logic within the static analysis engine 140 may be configured to de-obfuscate (e.g., decompile or disassemble) an object and obtain a higher level representation of the object than machine code, such as source code for example. Thereafter, the content of the source code is analyzed to determine if reflection operations would be invoked by the object when processed. After detection that the object would invoke reflection operations, the static analysis engine 140 provides the object under analysis or particular feature(s) associated with the object to the classification system for more in-depth analysis.

Additionally, or in the alternative, reflection can be identified by detecting function calls that invoke reflection operations, where the function calls may be directed to reflection APIs and/or system functions that invoke reflection operations. Hence, during processing of the object within the VM 163 1 and detecting at least one of the function calls that invoke reflection operations, the dynamic analysis engine 160 is able to determine that the object is suspicious.

Hereafter, the classification engine 180 is configured to determine whether an object, which is previously determined as suspicious, is further determined to be malicious or non-malicious. The object is deemed “suspicious” based on a determination of the presence of API calls within content of the object or a detection, during virtual processing of the object, of the issuance of function calls (e.g., API calls, system calls, etc.) that invokes reflection operations. The classification engine 180 may conduct probabilistic model analysis and/or machine learning analysis on certain feature(s) extracted from the object after a prior analysis uncovered that the object is invoking reflection operations. The feature(s) may include, but are not limited or restricted to function names, file sizes, and/or other information potentially indicative of malware such as extract suspicious strings from the contents of the object if the object has been successfully decompiled.

When implemented as hardware circuitry, the static analysis engine 140 may be configured to be communicatively coupled to communication interface logic 310 and/or the classification engine 180 . The dynamic analysis engine 160 may further be communicatively coupled to the communication interface logic 310 , the static analysis engine 140 , and/or the classification engine 180 . The classification engine 180 is communicatively coupled to the reporting engine 190 .

IV. Exemplary Threat Detection Based on Reflection

Referring to FIG. 4 , a general exemplary flowchart is shown that illustrates operations conducted by one or more electronic devices, such as a TDP or another type of platform, for determining whether a suspect object, which invokes reflection operations to obfuscate content or operability, is malicious. Upon receiving an object, an analysis is conducted to determine whether the suspect object is configured to access a reflection API (block 400 ). This may be determined by analyzing the de-obfuscated content associated with the object (e.g., the decompiled source code) for the presence of an API call that, at run-time, would invoke a reflection API. If the object includes such an API call, the object is deemed suspicious.

Additionally, or in the alternative, the behavior of the object may be monitored at run-time to detect whether the object is invoking reflection operations (block 410 ). For instance, this may be accomplished by setting interception points (e.g., hooks, breakpoints with subsequent activity after code execution halts, etc.) to detect one or more function calls resulting from processing the object within the virtual machine. One type of function call being monitored includes an API call directed to reflection API. Additionally, or in the alternative, another function call being monitored includes a system call that invokes reflection operations, where the system call may be based on an API call issued by the object.

In response to detecting that the object invokes reflection operations, content from the suspect object is extracted for further analysis (block 420 ). The content may include one or more features of the object under analysis, suspicious string data, or the like.

A classification analysis is conducted on the extracted content to determine the likelihood of the object, which invokes reflection operations, is associated with a malicious attack (block 430 ). According to one embodiment of the disclosure, the classification analysis may involve probabilistic model analysis and/or machine learning analysis to produce a result (e.g., a resultant score) that may be used to classify whether the object is malicious or not, as previously described. If the result is greater than a prescribed threshold, the suspect object is determined to be malicious (blocks 440 and 450 ). Otherwise, the suspect object is determined to be non-malicious (blocks 440 and 460 ).

Referring now to FIG. 5 , a first exemplary flowchart is shown that illustrates operations conducted by the static analysis engine and the classification system collectively deployed within the TDP and/or external resources (e.g. cloud services). Upon receiving a suspect object, an analysis is conducted to determine whether the object is configured to invoke reflection operations. This analysis may involve de-obfuscating by decompiling and/or disassembling (or by emulation) at least part of the object to recover a high-level representation (e.g., source code, or pseudo-code or another high-level language), and thereafter, conducting an analysis of the content that is part of the high-level representation (e.g., at least a portion of the source code or pseudo-code) to determine whether the object would invoke reflection at run-time (blocks 500 , 510 and 520 ). The object is considered to invoke reflection upon determining, by static analysis of the source code (or pseudo-code or another high-level language, that the code includes an API call to a reflection API.

If the de-obfuscated content of the suspect object fails to include an API call to a reflection API which is considered to be one of the triggering events for subsequent analysis, the analysis ends as the suspect object may be further analyzed through other malware detection schemes. However, in response to detecting that the suspect object is configured to access a reflection API for example, content from the suspect object is extracted for further analysis (blocks 520 and 530 ). The content may include one or more features of the suspect object (e.g., name of the reflection API, size of the suspect object, suspicious string data, or the like). Optionally, the static analysis engine may determine if the de-obfuscated (e.g., decompiled) high-level representation (e.g., source code, pseudo-code, or another high-level language) is further obfuscated, and if so, further operations are conducted to further de-obfuscate the high-level representation (blocks 540 and 550 ).

A classification analysis is conducted on the extracted content to determine the likelihood of the object being associated with a malicious attack (block 560 ). According to one embodiment of the disclosure, the classification analysis may involve probabilistic model analysis and/or machine learning analysis to produce a result that represents a likelihood of the object, which invokes reflection operations, is associated with a malicious attack, as previously described. If the result is greater than a prescribed threshold, the suspect object is determined to be malicious (blocks 570 and 580 ). Otherwise, the suspect object is determined to be non-malicious (blocks 570 and 590 ).

Referring to FIG. 6 , a second exemplary flowchart is shown that illustrates operations conducted by the dynamic analysis engine and the classification system collectively deployed within the TDP and/or external resources (e.g. cloud services). Upon processing the suspect object within a configured virtual machine, based on one or more behaviors of the object during processing within the virtual machine, a determination is made whether the object is invoking reflection operations (blocks 600 , 610 and 620 ).

In response to detecting that the object is invoking reflection operations, such as the object is attempting to access the reflection API for example, content from the object under analysis is extracted for further analysis (blocks 620 and 630 ). The content may include one or more features of the object, suspicious string data, or the like.

A classification analysis is conducted on the extracted content to determine the likelihood of the object being associated with a malicious attack (block 640 ). According to one embodiment of the disclosure, the classification analysis may involve probabilistic model analysis and/or machine learning analysis to produce a resultant score, as previously described. If the resultant score is greater than a prescribed threshold, the suspect object is determined to be malicious (blocks 650 and 660 ). Otherwise, the suspect object is determined to be non-malicious (blocks 650 and 670 ).

Referring now to FIG. 7 , an exemplary flowchart of the operations of the classification analysis described in FIGS. 4-6 as performed by the classification system of FIGS. 1 and 2 is shown. Herein, the classification system performs a first classification analysis on the content of the object to determine a first classification result (block 700 ). According to one embodiment of the disclosure, the first classification analysis includes a probabilistic model analysis on contents of the object, namely an analysis on features and other data associated with the object in accordance with a decision-tree analysis as described above. Based on these features provided for analysis, a result (e.g., resultant score) is produced, which represents the likelihood of the object under analysis being associated with a malicious attack. According to another embodiment of the disclosure, the first classification analysis may feature a machine learning analysis on content of the object, namely comparing content associated with the object to content associated with known malware or known benign data. Based on these comparisons, a result (e.g., resultant score) is produced, which represents a likelihood that the object is associated with a malicious attack.

Next, a determination is made whether additional classification analysis is to be performed (block 710 ). If so, the classification system performs a second classification analysis on the content of the object to determine a second classification result (block 720 ). Where the first classification analysis is directed to a probabilistic model analysis of content associated with the object, the second classification analysis may feature a more detailed probabilistic model analysis or a machine learning analysis. Similarly, where the first classification analysis includes a machine learning analysis, the second classification analysis may feature a more detailed machine learning analysis or a probabilistic model analysis.

In the foregoing description, the invention is described with reference to specific exemplary embodiments thereof. For instance, some or all of the functionality of the static analysis engine, the dynamic analysis engine and the classification engine of FIG. 1 may be implemented within another type of network device, such as an endpoint device. It will, however, be evident that various modifications and changes may be made thereto without departing from the broader spirit and scope of the invention as set forth in the appended claims.