Systems and methods for automated cybersecurity analysis of extracted binary string sets

Abstract

An automated system and method for analyzing a set of extracted strings from a binary is disclosed including processing the binary with a string-extraction logic that can locate strings within the binary and output an extracted string set for use in cybersecurity analysis. The logic retrieves a set of training data comprising a plurality of previously analyzed extracted string sets where each element of the previously analyzed extracted string set comprises at least one extracted string and a corresponding previously determined threat prediction score. A prediction model based upon the training data is generated and the extracted string set is processed by the prediction model to determine a threat prediction score for each string. Ranking of the located strings is based upon the determined threat prediction score, and an output of a ranked string list is generated.

Description

FIELD

Embodiments of the disclosure relate to the field of cybersecurity. More specifically, certain embodiments of the disclosure relate to a system, apparatus and method for an automated analysis of an extracted set of strings.

BACKGROUND

Over the last decade, malicious software (malware) has become a pervasive problem for Internet users and system administrators of networks devices. To counter this increasing problem, computer files are often inspected to verify that they do not contain any malware. Malware analysts, reverse engineers, forensic investigators, and incident responders have developed an arsenal of tools at their disposal to dissect malware and examine it for potential threats or other indications of source.

A “string” is a data type that comprises any finite sequence of characters (i.e., letters, numerals, symbols and punctuation marks). Data types are frequently used in programming languages as a way of categorizing data. Data types can differ according to the programming language used, however strings are implemented as a data type in virtually every programming language. The characters within strings are typically encoded in accordance with the American Standard Code for Information Interchange (ASCII) standard which establishes a relationship between the binary values stored within data and a pre-established set of characters. Other encodings and standards can be used to format strings including the Extended Binary Coded Decimal Interchange Code (EBCDIC) and UNICODE. Strings can be used for many purposes within computer files, including, for example encoding text relating to an error message that is displayed to the user upon triggering, a registry key, a uniform resource locator (URL) link, or a directory location for where to copy or store data within a computer system.

Malware analysis tools can examine strings contained within software binaries, namely any type of executable code including an application, script or any set of instructions. This examination may aide in gathering clues about the binary's function, threat level, design detection methods, and how containment of any potential damage may be achieved. For example, strings that contain filenames, internet protocol (IP) addresses, Uniform Resource Locators (URLs), domain names or the like may constitute indicators of compromise, and thus, are associated with a higher relevance to cybersecurity than strings that contain, for example, random sequences of characters. By analyzing suspicious binaries with a string extractor, a listing of the strings found within that binary can be generated. However, as the complexity of software and other binaries increases, the amount of strings to be reviewed as well as the effort required to determine relevance also increases. Hence, there is a need for a system to automatically locate and analyze sets of strings contained with various suspicious binaries under review.

BRIEF DESCRIPTION OF THE DRAWINGS

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. 1A depicts an exemplary system diagram of a cloud-based automated string analysis system in accordance with various embodiments of the invention.

FIG. 1B depicts an exemplary system diagram of a device-based automated string analysis system in accordance with an embodiment of the invention.

FIG. 1C depicts an exemplary hardware block diagram of an automated string analysis device in accordance with an embodiment of the invention.

FIG. 2 depicts an exemplary block diagram of an automated string analysis process in accordance with an embodiment of the invention.

FIG. 3 depicts an exemplary block diagram of automated prediction model generation utilizing string feature extraction in accordance with an embodiment of the invention.

FIG. 4A depicts an exemplary simplified list of extracted strings prior to automated analysis in accordance with an embodiment of the invention.

FIG. 4B depicts an exemplary simplified ranked list of extracted strings after automated analysis in accordance with an embodiment of the invention.

FIG. 5 depicts an exemplary flowchart of an automated process of extracted string set analysis in accordance with an embodiment of the invention.

DETAILED DESCRIPTION

Various embodiments of the disclosure relate to an automated system and/or process configured to analyze extracted string sets. This can be accomplished by generating a ranked list of extracted strings for use by various security systems and users within the cybersecurity field. According to one embodiment of the disclosure, the ranked list of extracted strings can be generated by ordering the extracted strings contained within a binary based on a generated threat detection score that corresponds to the likelihood of the string being associated with a risky or otherwise malicious action within the binary.

One of the tools malware analysts typically used when attempting to examine strings located within suspicious binaries is STRINGS.EXE from Sysinternals (a business unit of the Microsoft Corporation of Redmond, Wash.). STRINGS.EXE is an analytic software tool that is configured to receive a passed-in binary and scan it for embedded ASCII or UNICODE strings located within. However, certain analytic tools, such as STRINGS.EXE for example, may simply scan, extract, and generate an unordered list of strings that were located within the passed-in binary. No further analysis is done. By default, certain analytic tools, such as STRINGS.EXE identify strings as any sequence of characters comprising three or more consecutive characters followed by a null terminator. This type of indiscriminate string identification typically leads to the generation of noisy data sets since many of the extracted strings can be irrelevant, and thus obscure the highly relevant strings within the extracted string set.

For example, a set of consecutive bytes within a binary may be interpreted as a set of ASCII characters by STRINGS.EXE and thus be added to the list of extracted strings. However, the consecutive bytes may not actually represent a string of ASCII characters relevant for malware analysis, but instead represent irrelevant data such as a memory address, central processing unit (CPU) instruction, or other non-string data utilized within a program. As a result, string sets generated from analytic tools, such as STRINGS.EXE for example, often require human analysts to manually examine the extracted string sets in order to determine if relevant strings for malware analysis are present. The process of extracted string set analysis, which includes understanding and scoring the relevance of various extracted strings, along with manually generating a threat score often requires highly experienced human analysts. As a result, obtaining quality security-relevant scored data can be time consuming and expensive to obtain. Often, within an extracted string set, the frequency of relevant strings occurring within a set are disproportionately less than irrelevant strings. Additionally, during the manual analysis, variations may exist between the subjective opinions of various human analysts as to what strings constitute potential threats compared to other strings, based on differing past experiences or biases.

As the complexity of software and other binaries increases, the amount of time needed to manually analyze the extracted string sets also grows. Furthermore, as these extracted string sets grow in size, human error during the manual analysis process can also increase. For example, a human analyst may inadvertently skip over relevant strings during the manual review of the string set due to fatigue. Having an automated process to analyze extracted string sets may aide malware analysis by freeing up human analysts to examine other threat indicators of a suspicious binary under analysis.

Generating heuristic rules to robustly account for all possible variations of string combinations that may be extracted from suspicious binaries would be a monumental task. Thus, embodiments disclosed herein utilize automated machine learning frameworks to analyze extracted string sets and generate a ranked list output based on generated threat prediction scores. Many of the embodiments utilize an automated learning to rank (LTR) method to generate a potential threat score and utilize this score to create a ranking for each string extracted from a suspicious binary. LTR methods incorporate supervised machine learning procedures that utilize previously scored data to generate a prediction model that can then be used to predict a score for a new, previously unanalyzed data set (i.e. query) which can then be ranked based on the predicted score. Since the ranking (in many embodiments described herein) is related to the predicted threat scores of the extracted strings, the LTR ranking can be utilized to generate a ranked list of extracted strings. The rank list is an arrangement (i.e., ordering) of the strings according to rank. That is, a sequential arrangement based on the predicted scores of the extracted strings. For example, strings at the beginning of the ranked list can correspond to strings with a higher threat prediction score than subsequent strings in the list. As a result, rankings generated from the disclosed automated methods can be used to generate a ranked list of extracted strings which can subsequently be incorporated into a threat warning for further analysis or presentation to a threat detection system or human analyst.

LTR methods typically generate a prediction model function from known training data sets. The generated prediction model function can then receive new input data and output a score associated with the input data. The generation of a prediction model function is typically done by utilizing a large set of training data that have previously been analyzed and scored, often by a human analyst. Training data can be obtained from historical data generated from prior analyses. Once the prediction model has been generated, new data sets may be processed with the prediction model in order generate predictive scores for these new data sets without the need for human intervention. Embodiments herein utilize prediction models generated via automated machine learning methods to generate prediction scores associated with potential threat levels. By ranking strings extracted from a suspicious binary based on the predicted potential threat levels, the ranked strings at the beginning portion of the ranked string list are more likely to be relevant to further malware analysis compared to the strings in the later portion of the list.

Extracted strings can be expressed as feature data relating to features of the extracted strings. Feature data is typically represented as a number or other designation that correlates with a particular characteristic of the string. For example, a feature could be associated with the string that denotes the number of characters in the string, how many characters of a certain type are present, or if the string reads as natural language (denoting higher relevance) or as gibberish (denoting lower relevance). This type of feature extraction can be accomplished by utilizing natural language processing tools. In this way, some embodiments may generate a machine learning prediction model that utilizes feature data to further minimize the influence of irrelevant strings or random sequences of characters not probative of a cyberattack or otherwise meaningful to cybersecurity. Additionally, in certain embodiments, the automated LTR prediction model may utilize similar string feature data comprised within training data to create a prediction model that can analyze an extracted string against historical threat prediction scores, along with string feature data to generate threat prediction scores with increased accuracy.

In certain embodiments, the generated prediction model may utilize a gradient boosted decision tree (GBDT) method for the machine learning prediction model. LTR systems can be understood as a pairwise classification system, meaning that the system evaluates pairs of items (e.g., extracted strings) from a set at a time and iteratively computes the optimal ranking for all pairs of items (e.g., extracted strings) within a set to come up with a final ranking for the entire set. GBDT methods generally incorporate individual decision trees to facilitate prediction score generation by using a weighted sum of the leaves of each decision tree. GBDT methods can classify each pair of extracted strings as correctly or incorrectly ranked, and use the optimal ordering of each pair of extracted strings to come up with the final ranking for all of the extracted strings within the extracted string set.

Once a suspicious binary has been fully processed and the associated extracted strings ranked with corresponding predicted threat values, a threat warning can be generated. Threat warnings may utilize predetermined rule sets or thresholds to process the ranked extracted string set. In some embodiments, the suspicious binary is initially analyzed in response to a user request (where the “user” may be, e.g., a computer user, security analyst or system admin) and the threat warning is then utilized for the generation of a threat report that is presented to the user. In other embodiments, in response to a predicted threat level that is beyond a predetermined threshold, the threat warning may be utilized to create a remedial action. In certain cases, a score for a single string (e.g., a reference to a particular, known sensitive memory address, etc.) may be enough to generate a remedial action on the entire binary.

Threat warnings can be utilized to generate threat reports, emails, or other communications presented (e.g., displayed or sent) to a user informing them of the results of the analyzed extracted strings within the suspicious binary. These threat warnings can be utilized, in various embodiments, to auto-generate the threat report, email, or other communication informing the user of the results of various analyses. In certain embodiments, the threat warnings can also be reported to outside third parties.

Threat warnings may also be utilized to update remedial action behaviors as responses to newly determined threats as they are identified, such as, but not limited to, a new malware attack having a particularly new threat pattern. In some embodiments, remedial actions may be taken without human intervention. In further embodiments, the string analysis logic may be given a set of pre-defined thresholds and/or rules that may empower the generated threat warnings to initiate remedial actions immediately based on the predicted threat data derived from the suspicious binary. Remedial actions may include, but are not limited to, quarantining the suspicious binary within a system, or halting any processing of the binary.

It should be understood that threat warnings are utilized by varying types of users with unique needs. For example, malware analysts are typically skilled at reading strings. Such malware analysts typically examine string threats as leads for deeper analysis of the suspicious binary which can lead to a variety of outcomes including classification of the malware, verification that the suspicious binary is in fact malware, or mapping of the malware to a certain family. Security operations center (SOC) analysts generally respond to alerts coming into the system. These alerts are typically examined to determine if a suspicious binary needs to be escalated for further review. SOC analysts can benefit from concise, pre-generated threat reports incorporating ranked string lists that can focus their attention on relevant strings. Specifically, utilizing the methods and systems described herein can reduce the time needed for intervention by a SOC analyst. Incident response (IR) consultants conduct investigations of specific incidents of intrusions and other cyberattacks in progress or having occurred, and how to remediate them. For these users, having a focused ranked set of strings can help determine where the malware may be on other areas or systems within the network. Finally, threat intelligence analysts typically want to know if a piece of discovered malware is related to other known pieces of malware. By looking at a ranked string list, the threat intelligence analyst may be able to see similar strings as those found in previous malware, which can help indicate similar origins.

Another aspect of the invention is that the resulting prediction models that are generated for predicting threats on new suspicious binaries can be analyzed, verified, and compared quantitatively. By using such quantitative methods, the prediction model's performance can be assessed and given a value to compare to other models. In some embodiments, the quantitative method utilized is a mean normalized discounted cumulative gain (MNDCG) method that generates a score of each item within a generated prediction model. Broadly, this method examines the magnitude of each string's relevance summed over the entire string set, which can be represented as a non-negative number called the “cumulative gain”. The MNDCG method can then discount these results within the prediction model in a typically logarithmic fashion so as to reflect the goal of having the most relevant strings appear, for example, towards the top of the predicted ranking. That output is normalized so the results of the MNDCG method can be compared to other generated prediction models of varying size. Finally, the quantitative evaluation can limit a certain number of strings of a binary from appearing at (or near, e.g., within a predetermined distance (i.e., number of strings) from) the top of the ranked string list to help limit the computational requirements needed (as some suspicious binaries may have thousands or tens of thousands of binaries). For example, the quantitative analysis results may limit the output to the first 100 strings of a binary within a ranked string list, but could be adjusted via a user interface by an analyst examining the suspicious binary.

It is understood that the process described herein provide for a more efficient and robust method of providing ranked strings sets for malware analysis in an automated fashion. The automated generation of threat predictions scores on new, previously unranked string sets which can be utilized to rank and generate threat warnings can provide a more accurate threat assessment of suspicious binaries as well as increasing efficiency through reducing the time needed for a human analyst to review the set. This facilitates the practical application of providing more efficient malware detection.

I. Terminology

In the following description, certain terminology is used to describe features of the invention. For example, in certain situations, the term “logic” is representative of hardware, firmware or software that is configured to perform one or more functions. As hardware, logic may include circuitry such as one or more processors (e.g., a microprocessor, one or more processor cores, a virtual central processing unit, a programmable gate array, a microcontroller, an application specific integrated circuit, etc.), wireless receiver, transmitter and/or transceiver circuitry, semiconductor memory, combinatorial logic, or other types of electronic components.

As software, logic (or “engines” in certain descriptions) 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/dynamic load library, or one or more instructions. These software modules may be stored in any type of a suitable non-transitory storage medium, cloud-based 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 non-transitory storage mediums may include, but are not limited or restricted to a programmable circuit; a semiconductor memory; 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. As firmware, the executable code is stored in persistent storage.

The term “malware” is directed to software that produces an undesirable behavior upon execution, where the behavior is deemed to be “undesirable” based on customer-specific rules, manufacturer-based rules, or any other type of rules formulated by public opinion or a particular governmental or commercial entity. This undesired behavior may include a communication-based anomaly or an execution-based anomaly that (1) alters the functionality of a network device executing application software in a malicious manner; (2) alters the functionality of a network device executing application software without any malicious intent; and/or (3) provides an unwanted functionality which is generally acceptable in other contexts.

The term “object” generally refers to content in the form of an item of information having a logical structure or organization that enables it to be classified for purposes of analysis for malware. One example of the object may include an email message or a portion of the email message. Another example of the object may include a storage file or a document such as a PHP or other dynamic file, a word processing document such as Word® document, or other information that may be subjected to cybersecurity analysis. The object may also include an executable such as an application, program, code segment, a script, dynamic link library “dll,” URL link, or any other element having a format that can be directly executed or interpreted by logic within the network device. Network content such as webpages and other downloaded content may be further examples of objects analyzed for malware.

The term “binary” embraces a computer program code that represents text, computer processor instructions, or any other data using a two-symbol system, such as, for example, “0” and “1” from the binary number system. A binary code assigns a pattern of binary digits, also known as bits, to each instruction. Binary codes are used to encode data, such as each digit or character, into bit strings of fixed-width or variable width, depending on the implementation. The term “binary”, as used herein, may also designate an executable or interpretable computer processor instruction, regardless of whether in a two-symbol system, depending on context of its use in this description. For example, a “binary” may refer to any non-text file, but which may nevertheless comprise embedded text as strings. One example of a non-text file that may contain embedded strings is a text-editor file that comprises not only the text within the document, but also includes data related to formatting the text within the program. Binaries may include a variety of types of objects, such as executables, applications, programs, scripts, etc. It is understood that the term binary may include partial, corrupt, or otherwise incomplete files.

The term “cloud-based” generally refers to a hosted service that is remotely located from a data source and configured to receive, store and process data delivered by the data source over a network, including a self-hosted and third-party hosted service. Cloud-based systems may be configured to operate as a public cloud-based service, a private cloud-based service or a hybrid cloud-based service. A “public cloud-based service” may include a third-party provider that supplies one or more servers to host multi-tenant services. Examples of a public cloud-based service include Amazon Web Services® (AWS®), Microsoft® Azure™, and Google® Compute Engine™ as examples. In contrast, a “private” cloud-based service may include one or more servers that host services provided to a single subscriber (enterprise) and a hybrid cloud-based service may be a combination of both a public cloud-based service and a private cloud-based service.

The term “network device” should be generally construed as electronics with data processing capability and/or a capability of connecting to any type of network, such as a public network (e.g., Internet), a private network (e.g., a wireless data telecommunication network, a local area network “LAN”, etc.), or a combination of networks. Examples of a network device may include, but are not limited or restricted to, the following: a server or other stand-alone electronic device, a mainframe, a firewall, a router; an info-entertainment device, industrial controllers, vehicles, or a client device (e.g., a laptop, a smartphone, a tablet, a desktop computer, a netbook, gaming console, a medical device, or any general-purpose or special-purpose, user-controlled electronic device).

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, steps or acts 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 not intended to limit the invention to the specific embodiments shown and described.

II. System Architecture

Referring to FIG. 1A, an exemplary system diagram of a cloud-based automated string analysis system 100A is shown. The string analysis can be accomplished within a private virtual cloud system 135, which, in an embodiment, may be provided within a larger public cloud service 130A. In many embodiments, a client device 140A communicatively coupled to the public cloud service 130A via a network 120 can receive a suspicious binary from a source device 110 which is also communicatively coupled to the network 120. In response to receiving a suspicious binary, the client device 140A can send the binary to the public cloud service 130A which forwards it to the private virtual cloud 135.

The private virtual cloud 135 and associated resources can be generated as part of an Infrastructure-as-a-Service (IaaS) model and comprise at least one instance of a vCPU 136 and a memory 137 communicatively coupled to the vCPU 136. It would be understood by those skilled in the art that the private virtual cloud 135 may comprise a variable number of vCPUs and memory stores as needed based on various factors including, but not limited to, the current available computing resources available on the system, or the current computational demands placed upon the private virtual cloud 135.

The memory 137 can have string analysis logic to direct the vCPU 136 to process the received suspicious binary. In response, the vCPU 136 extracts the strings from the suspicious binary, generates a prediction model for evaluating the extracted strings, and processes the suspicious binary through the prediction model function to generate a list of prediction scores that can be utilized to generate a ranked list of strings taken from the suspicious binary that correlate to the perceived threat of each string. The ranked list of extracted strings can then be utilized to generate an overall threat warning for the suspicious binary. In certain embodiments, the ranked list of extracted strings may be sent back to the client device 140A for further processing. In various embodiments, the string analysis of the suspicious binary may result in a determination that the binary poses an immediate threat such that remedial action should be taken, which may then occur or be communicated to the client device 140A for further action. In further embodiments, the resulting ranked string list may be sent to a third party, such as cybersecurity vendors, or other external threat analysts for evaluation in a threat report. In certain embodiments, the memory 137 comprises prediction model generation logic that may utilize public data, non-public data, or a mixture of both to generate a prediction model function that can be accessed or otherwise provided to a client device 140A for supplementing string analysis logic within the client device 140A. In certain embodiments, the prediction model may be sent to an analyst station or admin for continued action such as further threat analysis and/or remediation.

In some embodiments, the private virtual cloud 135 may be accessed by a network security appliance 150 which may require assistance in evaluating the threat of a suspicious binary received from a source device 110. Embodiments relating to the network security appliance 150 can behave similarly to the embodiments of the client device 140A such that communication between the network security appliance 150 and private virtual cloud 135 is analogous to the communication between the client device 140A and private cloud server 135.

Referring to FIG. 1B, a system diagram of a device-based automated string analysis system 100B is shown. The string analysis device 130B can be located within a client network 160 and be in communication with at least one client device 140B. As the client network 160 receives a suspicious binary from the source device 110 over a network 120, the client device 140B can pass the binary to the string analysis device 130B for processing.

Similar to the embodiments discussed above with respect to FIG. 1A, communication between the client device 140B and the string analysis device 130B may be analogous to the above description of the communication between the private virtual cloud 135 and the client device 140A of FIG. 1A. As would be understood by those skilled in the art, the client network 160 may comprise any number of devices, including multiple client or network devices. Additionally, the string analysis device 130B may also operate as a subsystem of a larger network security appliance within the client network 160. In further embodiments, the string analysis device 130B may be implemented as a virtual instance (i.e. software) that runs on the client device 140B.

Referring to FIG. 1C, a hardware block diagram of an automated string analysis device 130B is shown. The string analysis device 130B comprises a network interface 131 which can be utilized to connect to the network 120 for communication similar to the discussion of FIG. 1A. The string analysis device 130B further comprises a processor 132, memory 170, and training data store 133 which are all in communication with each other. The memory further comprises a plurality of logics including, but not limited to, string extraction logic 171, prediction model generation logic 172, prediction model verification logic 173, ranking logic 174, and reporting logic 175.

The string analysis device 130B can receive suspicious binaries from the network 120 via the network interface 131. In response, the processor 132 can be instructed by the string extraction logic 171 to extract the strings found within the suspicious binary. Once extracted, the prediction model generation logic 172 can utilize training data within the training data store 133 to generate a prediction model. It is contemplated that various embodiments of the string analysis machine 130B may utilize the prediction model generation logic 172 to retrieve a pre-generated prediction model from an external source over the network 120 instead of generating a new model internally. In fact, certain embodiments of the string analysis device 130B may not comprise a training data store 133 and instead can retrieve data (if needed) via a remote connection over the network 120.

Once generated or retrieved, the prediction model can be verified via the prediction model verification logic 173. The verification may be accomplished using verification data either extracted or derived from the training data, or via a specialized set of verification data that may be stored on the training data store 133 or on a remote device. The ranking logic 174 can be utilized to rank a set of strings based on the prediction model either during the verification process or during general analysis of strings extracted from suspicious binaries. The output of the ranking logic 174 can be analyzed by the reporting logic 175 to determine if a report should be generated, and if so, what actions to take. For example, the ranking of a set of strings may require that a threat report should be sent to an analyst for further evaluation. Ranked string sets may also be determined to contain strings that trigger a pre-determined rule (such as strings that are directed to specific, crucial locations within memory) that require the reporting logic 175 to trigger at least one remedial action, sometimes independently without human intervention. As those in the art will understand, the reporting logic 175 can be configured to generate and respond to a number of various threats in a variety of ways that can minimize the potential threat of malware determined to be contained within the suspicious binary.

It should be understood that although certain embodiments are highlighted in discussion of FIGS. 1A-1C, a wider variety of embodiments are possible and contemplated by this application. In fact, based on the desired application and layout, a mixture of client devices, source devices, and other components can be utilized in order to provide an automated system to analyze extracted strings.

III. Training and Prediction

Referring to FIG. 2, an exemplary block diagram of an automated string analysis process 200 is shown. Broadly, the process of string analysis can be understood to comprise two phases: training and predicting. Before predictions can be made, a prediction model 250 must be generated in the training phase 210. In order to better visualize the string analysis process 200, the elements associated with the training phase 210 are bounded by a dashed line. Training data binaries 220 are gathered and processed through a string extractor 230. The output of the string extractor 230 comprises a set of strings which are then, in many embodiments, evaluated by human analysts for relevance related to potential malware threats.

A plurality of these analyzed string sets 241-243 are conceptually shown in FIG. 2 as lists comprising rows of two associated elements. The left “X” element represents a string extracted from the training data binaries 220, while the right “Y” element corresponds to a score assigned to the string by a human analyst. Scores are often assigned as a non-negative integer number that corresponds to the string's relevance for malware analysis. In some embodiments, additional elements may be present in the analyzed string sets 241-243 that represent values associated with features of the string.

Each element in the analyzed string sets 241-243 comprises both a superscript and a subscript. The superscript denotes the rank of the element in the binary so the first row has a superscript of 1, the second row has a superscript of 2, and so on until the last row u is reached. Each analyzed string set 241-243 may have variable lengths which are denoted by the variables u, v, and w. The subscript denotes the number of the training files within the training data 240. Therefore, the first analyzed string set 241 has a subscript of 1 on every element, the second analyzed string set 242 has a subscript of 2 on every element, up to the last analyzed string set 243 which has a subscript of m denoting that the number of analyzed string sets within the training data 240 can be variable and may include large numbers of sets. In fact, in order to increase the robustness of the prediction model 250 generated from the training data 240, the training data 240 may include a large number of previously analyzed and ranked string sets (“labelled” string sets).

With the training data 240 comprising a plurality of analyzed string sets 241-243 generated from the training data binaries 220, the string analysis process 200 can generate a prediction model 250 that can be utilized to create prediction scores on subsequent binary strings. In many embodiments, prior to predicting subsequent strings, the prediction model 250 can undergo a verification process. Verification can occur in many ways, but may be accomplished by directing the prediction model to process at least one (but likely many) verification data sets. A verification data set can be a set of unranked strings which have previously been ranked and analyzed manually. Upon processing of the verification data set by the prediction model 250, the system 200 or analyst may compare the sorted data set 270 generated from the verification data set to the known ranking of the verification data set. In this way, the prediction model 250 can be verified for accuracy and, in certain embodiments, may be adjusted based on the processing of the verification data set. In further embodiments, the verification process may occur automatically based on a set of pre-determined heuristics or thresholds.

The subsequent binary strings for processing by the prediction model 250 may be pre-ranked string sets used to verify the prediction model 250 for accuracy in ranking (as described above), or unranked string sets encountered for example in a private network and potentially representing cybersecurity threats. The embodiment shown in FIG. 2 illustrates decision trees similar to GBDT methods that create an ordered ranking of the analyzed string sets 241-243 within the training data 240. Once the ranking is compete and the prediction model has been generated, the training phase is complete and moves into the prediction phase.

A suspicious binary 221 can be processed through the string extractor 230 to generate a new list of strings, commonly called a query 260. As denoted in FIG. 2, the query 260 has a superscript to denote the initial ordering of the strings (typically the order in which they were discovered during the string extraction). The number of elements in the query 260 can vary based on the size of the suspicious binary 221 and does not have to be similar in size to the analyzed string sets 241-243 in the training data 240. When the query 260 has been fully generated, it can be passed to the prediction model 250 for processing and prediction. The prediction model 250 typically takes each element of the query 260 and assigns a score (denoted as Ŷ) that is predicted for the string.

Once each element of the query 260 has been processed and received a score from the prediction model 250, the string analysis process 200 can perform a ranking of the elements based on the predicted scores for each element. The final ranked string list 270 comprises rows of a string element “X” and a predicted score element Ŷ. The superscript on each element of the row denotes its ranking within the ranked string list 270. Since the score elements correspond to the string's relevance for malware analysis, it can be understood that strings with a higher ranking will have more relevance for further analysis than strings with a lower ranking. The ranked string list 270 can then be utilized to create a threat warning which may be utilized to generate a threat report or other remedial action in response to the presence of a predicted score higher than a pre-determined threshold. The generation of prediction models utilizing string features is discussed in more detail below.

Referring to FIG. 3, an exemplary block diagram of automated prediction model generation utilizing string feature extraction is shown. Typically, a machine learning system 350 will access a set of training data 240 to begin the process of generating a prediction model 250. As discussed above, the training data 240 typically consists of a series of extracted string sets that have been previously analyzed and scored by human analysts or automated heuristics. The training data 240 is typically comprised of pairs of extracted strings and associated threat scores. The number of extracted strings within the training data 240 can be of any size, including, but not limited to, tens of millions of strings. Large amounts of training data can often lead to more accurate predictions. For example, as a training data set increases in size, biases within the human analysts who scored the training data are often diminished. As more data points associated with the training data are obtained or utilized within the prediction model 250, the overall error in the predicted score of a new query will generally be reduced. Data points may be expressed through the use of string features.

In many embodiments, the machine learning system 350 can process the strings associated with the training data 240, and extract features from the strings via a feature extractor 340. In other embodiments, the training data 240 already comprises feature data associated with each string. The machine learning system 350 can then generate a prediction model 250 which can accept a new query with extracted strings and associated string features and generate a predicted threat score and ranking to associate with that string.

Once received, a suspicious binary 221 can be analyzed by first extracting the strings within the suspicious binary 221 via the string extractor 230. Each located string within the suspicious binary 221 can then utilize a strings feature extractor 340 to determine various features 341-343. It should be understood that the number of features extracted can vary depending on the needs of the application. For example, fewer features can be extracted and utilized when computational resources or time is limited. However, more features may be utilized if the extracted string set is very large and increased differentiation between the string threat levels is needed.

By way of example and not limitation, a first feature 341 may be the length of the string. Strings of increased length may correspond to varying levels of threats. A second feature 342 can relate to the type of string. Strings that have been determined to contain natural language elements may be given a certain value compared to strings comprising random characters. String features can be derived from any meaningful distinction that varies between strings and that can be assigned a numerical value for comparison within the prediction model 250. Once all features 341-343 have been extracted from the set of strings located within the suspicious binary 220, the resulting query can be passed to the machine learning system 350 for processing within the prediction model 250 which generates a prediction score for each string. These generated prediction scores can be utilized to create a ranked string list for use in generating threat warnings. A simplified example of a ranked string list that can be generated with this method is discussed below.

Referring to FIG. 4A, an exemplary simplified list of extracted strings 400A prior to automated analysis is shown. The list of extracted strings 400A omits showing all rankings in order to reduce complexity, but it is understood that every slot position between the first slot and the last slot corresponds to an extracted string. In the list of extracted strings 400A, there are various strings that can be classified as having high, mid, or low relevance while some strings are classified as being irrelevant in regards to malware analysis. It is understood that the placement of each string in an unranked set of extracted strings generally correspond to the order in which the strings were located and processed by the string extractor. It should also be understood that the determination of relevance is typically not known at this point prior to analysis and ranking, but is present in FIG. 4A to highlight the processing of the list from an unranked to the ranked state shown in FIG. 4B.

In this example, the first slot in the list of extracted strings 400A corresponds to a first string that has a mid-relevance 410, the second slot corresponds to an irrelevant string 420, and the third slot corresponds to a high relevance string 430. Lower in the list, the fifteenth slot of the list of extracted strings 400A corresponds to a first low relevance string 440. Further down, the twenty third slot corresponds to another high relevance string 450, while the last slot corresponds to a third high relevance string 460. It should be understood that the length of the list of extracted strings 400A can be of any length depending on the number of strings located within a given suspicious binary. The list of extracted strings 400A can be processed by the string analysis logic to create a ranked list of extracted strings such as the one in FIG. 4B.

Referring to FIG. 4B, an exemplary simplified ranked list of extracted strings after automated analysis is shown. Similar to the list of extracted strings 400A of FIG. 4A, the ranked list of extracted strings 400B omits showing all rankings in order to reduce complexity, but it is understood that every ranked slot between the first slot and the last slot corresponds to a string. The ranked list of extracted strings 400B is further understood to be comprised of the same strings as the list of extracted strings 400A prior to automated analysis, and that these strings have been sorted to different slots based upon their relevance as determined by their associated threat prediction scores.

As can be seen in the ranked list of extracted strings after automated analysis 400B, the ranking of the given strings corresponds to a descending order from high relevant strings, down to mid relevant strings, then to low relevant strings, and then concluding with irrelevant strings. The first three ranked slots correspond to the first, second and third high relevance strings 430, 450, 480. The fifteenth slot corresponds to the mid relevance string 410 which was originally located first in the unranked list of extracted strings 400A. The twenty-third slot corresponds to the low relevance string 440, while the last ranked slot in the ranked list of extracted strings 400B corresponds to the irrelevant string 420 that was previously located in the second slot of the list of extracted strings 400A.

As can be understood, the length of the ranked list of extracted strings 400B can be any length and typically corresponds both to the number of strings extracted from the suspicious binary, but also to the number of ranked slots in the list of extracted strings 400A. However, it is contemplated that certain embodiments may generate a ranked list of extracted strings 400B that is smaller than the unranked list of extracted strings 400A due to the process of eliminating strings that are below a certain relevance threshold or ones that are duplicative. Once generated, the ranked list of extracted strings 400B can then be utilized to generate a threat warning suitable for further malware analysis.

IV. String Analysis Process

Referring now to FIG. 5, a flowchart of the automated extracted string analysis process 500 is shown. The process 500 typically begins with the receiving of a suspicious binary (block 510). In many embodiments, this binary can be received from a source device, which could be a hardware-based source, a virtualized source, or any source that is communicatively coupled with the string analysis logic. However, as described above, the source of the received binary can often be from the client device or network security device that is evaluating the binary for malware.

Upon reception of a suspicious binary, the process 500 can utilize a tool to extract the strings from the suspicious binary (block 520). The tool utilized may be, for example, STRINGS.EXE, but may instead be any suitable analytic tool that can generate a list of strings located within a given binary. In a number of embodiments, the generated list of strings is further processed to extract features from the strings (block 530). Features are often processed as a numerical value corresponding to a certain property of the string. As discussed above, features may be related to various aspects of the string including, but not limited to, length, type, frequency, or similarity to previous features. Upon extracting the features from the generated list of strings, the data is ready for processing within a prediction model.

The process 500 often requires that a set of training data be retrieved from a training data source (block 540). In many embodiments, the training data is stored in a memory communicatively coupled to the string analysis logic and can be retrieved directly. In other embodiments, the training data may be available through a remote service and must be requested to be retrieved. Once the training data has been retrieved, a prediction model may be generated based on the training data (block 550). The prediction model can typically be logic that receives new extracted strings and generates a prediction score based on the training data used to generate the prediction model. Often, this can be achieved through the use of a GBDT method. Because the training data comprises scores relating to threat levels associated with previously analyzed extracted strings, the prediction score generated by the prediction model corresponds to a prediction for the perceived level of threat of the new extracted strings.

Once the prediction model is generated, it can then be verified through the use of verification data (block 555). Typically, the verification can be accomplished by running a set of previously verified string data through the prediction model and comparing the results of the prediction model to the results previously determined on the same string data. As a result of the verification process, the prediction model may be adjusted which can be accomplished either manually or via an automated adjustment process based on a set of pre-established thresholds or heuristics. Once verified, the string analysis ranking logic can process each string from within the extracted string set to generate a prediction score that is associated with each analyzed string (block 560). In certain embodiments, the prediction score is a non-negative value that can range from within a lower and upper bound. For example, the threat levels could be assigned a number between zero and seven, with seven being a higher predicted threat than a string associated with a zero prediction score.

Upon completion of the prediction score generation, the process 500 can generate a new string list that ranks the order of the analyzed extracted strings based upon the threat prediction score (block 570). In certain embodiments, the process of generating the ranked list of strings may delete duplicate strings, or add an indicator within the list to each string to denote the number of occurrences of the extracted string within the ranked set. In this way, the ranked list of strings can be displayed in a more efficient manner. In various embodiments, the process of generating the ranked list of extracted strings may limit the number of strings within the ranked set based on factors including, but not limited to, the total number of extracted strings within the original set, or strings that only exceed a certain threat threshold value level. For example, a ranked list of extracted strings may only comprise the first one-hundred entries even if the binary under analysis comprises thousands of strings. The generated ranked list of extracted strings may then be utilized in various ways.

The process 500 can utilize the ranked list of extracted strings to generate an overall threat warning (block 580). As discussed above, threat warnings can have a variety of uses such as generating remedial actions, or providing data for the generation of an overall threat report. Based upon the requirements of the user, the presence of strings that exceed a pre-determined value can create a trigger for a network security system to take immediate remedial action such as, but not limited to, quarantining the suspicious binary within a system, or to halt processing of the binary.

Threat reports may incorporate the threat warning data into the threat report in order to aid a human analyst's further threat analysis of the suspicious binary (block 590). As discussed above, the type of data and report required by the analyst can vary depending on the nature of the analysis sought. For example, the string rankings can be used in an analyst's further investigation and analysis of the cybersecurity threat represented by the binary through analysis of the order of the strings within the ranked strings. The binary analysis results may further be combined with other cybersecurity information and indicators of compromise to aid in determining whether a cyberattack is occurring and/or the remediation appropriate to mitigate the attack and its damage.

The binary utilized in the string ranking analysis may also be subjected to further static and/or dynamic analysis. These types of analysis employ a two-phase malware detection approach to detect malware contained in suspicious binaries or other network traffic monitored in real-time. In a first or “static” phase, a heuristic is applied to a suspicious binary or other object that exhibits characteristics associated with malware. In a second or “dynamic” phase, the suspicious objects are processed within one or more virtual machines and in accordance with a specific version of an application or multiple versions of that application associated with the binary. These methods offer a two-phase, malware detection solution with options for concurrent processing of two or more versions of an application in order to achieve significant reduction of false positives while limiting time for analysis. Static and dynamic analysis techniques that can be utilized in accordance with embodiments of the invention are described in U.S. Pat. No. 9,241,010 issued Jan. 19, 2016 and U.S. Pat. No. 10,284,575, issued May 7, 2019, the disclosures of which are hereby incorporated by reference in their entirety.

In the foregoing description, the invention is described with reference to specific embodiments thereof. 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.

Claims

1. An automated computerized method for analyzing a set of extracted strings relevant for cybersecurity threat detection comprising: processing a binary with a string-extraction logic, wherein the string extraction logic is configured to locate strings within a received binary and output an extracted string set of the located strings;processing the extracted string set with a prediction model generated from a set of training data to determine a threat prediction score for each located string within the extracted string set;ranking the located strings within the extracted string set based upon the determined threat prediction score; andoutputting a ranked string list based upon the located strings' ranking,wherein prior to the processing of the extracted string set, the prediction model is generated based on at least the set of training data including a plurality of previously analyzed extracted string sets, each element of the previously analyzed extracted string sets comprises at least one extracted string and a corresponding previously determined threat prediction score.

2. The method of claim 1, wherein the located strings associated with a higher threat prediction score appear in the ranked string list before strings associated with a lower threat prediction score.

3. The method of claim 1, wherein in response to the outputting of the ranked string list, generating a threat warning comprising additional cybersecurity threat data associated with the ranked string list.

4. The method of claim 3, wherein, in response to the threat warning exceeding a first pre-determined threshold, generating a threat report incorporating the ranked string list.

5. The method of claim 4, wherein the threat report only incorporates strings from the ranked string list that exceed a second pre-determined threshold.

6. The method of claim 4, wherein the set of strings incorporated within the threat report does not comprise duplicate strings.

7. The method of claim 3, wherein in response to the threat warning exceeding a pre-determined threshold, a remedial action is conducted.

8. The method of claim 1, wherein the method is practiced at least partially within a cloud-based computing environment.

9. An automated computerized method for analyzing a set of extracted strings relevant for cybersecurity threat detection comprising: processing a binary with a string extraction logic, wherein the string extraction logic is configured to locate strings within a received binary and output an extracted string set of the located strings;processing the extracted string set with a prediction model generated from a set of training data to determine a threat prediction score for each located string within the extracted string set;ranking the located strings within the extracted string set based upon the determined threat prediction score; andoutputting a ranked string list based upon the rankings of the located strings,wherein the prediction model utilized to generate the ranked string list is further processed within a quantitative analysis system to generate a first comparative score suitable for comparison with a second comparative score associated with a second prediction model utilized to generate a second ranked string list in order to assess the validity of the prediction model utilized to generate the ranked string list.

10. The method of claim 9, wherein the quantitative analysis system utilizes normalized discounted cumulative gain methods.

11. An automated system for analyzing a set of extracted strings relevant for cybersecurity threat detection comprising: a processor; anda transitory storage medium communicatively coupled to the processor, the transitory storage medium includes string analysis logic configured to: process a binary with a string extraction logic, wherein the string extraction logic is configured to locate strings within a received binary and output an extracted string set of the located strings;process the extracted string set with a prediction model to determine a threat prediction score for each located string within the extracted string set;rank the located strings within the extracted string set based upon the determined threat prediction score; andoutput a ranked string list based upon the ranking of the located strings,wherein the prediction model is generated, prior to the processing of the extracted string set based on at least a set of training data comprising a plurality of previously analyzed extracted string sets andwherein each element of the previously analyzed extracted string sets comprises at least one extracted string and a corresponding previously determined threat prediction score.

12. The system of claim 11, wherein the located strings associated with a higher threat prediction score appear in the ranked string list before strings associated with a lower threat prediction score.

13. The system of claim 11, wherein in response to the outputting of the ranked string list, a threat warning is generated comprising additional cybersecurity threat data associated with the ranked string list.

14. The system of claim 13, wherein a threat report incorporates the ranked string list and is generated in response to the threat warning exceeding a first pre-determined threshold.

15. The system of claim 14, wherein the threat report only incorporates strings from the ranked string list that exceed a second pre-determined threshold.

16. The system of claim 14, wherein the ranked string list incorporated within the threat report does not comprise duplicate strings.

17. The system of claim 15, wherein remedial action is taken in response to the threat warning exceeding a third pre-determined threshold.

18. The system of claim 11, wherein the prediction model utilized to generate the ranked string list is further processed within a quantitative analysis system to generate a first comparative score suitable for comparison with a second comparative score associated with a second prediction model utilized to generate a second ranked string list in order to assess the validity of the prediction model utilized to generate the ranked string list.

19. The system of claim 18, wherein the quantitative analysis system utilizes normalized discounted cumulative gain.

20. The system of claim 11, wherein the system is at least partially operated within a cloud-based computing environment.

21. The system of claim 20, wherein the processor is operated within a virtual computing environment.

22. An automated system for analyzing a set of extracted strings relevant for cybersecurity threat detection comprising: a processor; anda transitory storage medium communicatively coupled to the processor, the transitory storage medium comprises: a string extraction logic to process a binary to locate strings within the binary and output an extracted string set of the located strings;a prediction model logic configured to retrieve a prediction model generated with a set of training data and verified with a set of verification data;a ranking logic configured to rank the located strings within the extracted string set based on a prediction score generated by the prediction model for each located string; anda reporting logic configured to generate a threat warning comprising data generated from the ranked string list wherein the threat warning is formatted for a human analyst to perform further analysis.

23. The automated system of claim 22, wherein the located strings associated with a higher threat prediction score appear in the ranked string list before strings associated with a lower threat prediction score.

24. The automated system of claim 22, wherein the reporting logic to generate the threat warning comprising additional cybersecurity threat data associated with the ranked string list in response to an outputting of the ranked string list.

25. The automated system of claim 22, wherein responsive to the threat warning exceeding a pre-determined threshold, generating the threat report incorporating the ranked string list.