Method and system for detecting malware

Abstract

A system and method of analysis. NX domain names are collected from an asset in a real network. The NX domain names are domain names that are not registered. The real network NX domain names are utilized to create testing vectors. The testing vectors are classified as benign vectors or malicious vectors based on training vectors. The asset is then classified as infected if the NX testing vector created from the real network NX domain names is classified as a malicious vector.

Description

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 illustrates a system for detecting malware, according to one embodiment.

FIGS. 2-4 illustrate a method for detecting malware, according to one embodiment.

FIG. 5 illustrates various elements involved in domain name resolution.

FIGS. 6-10 illustrate examples for detecting malware, according to several embodiments.

DESCRIPTION OF EMBODIMENTS OF THE INVENTION

FIG. 1 illustrates a system for detecting malware, according to one embodiment. FIG. 1 illustrates at least one network 101 (e.g., the Internet) connecting at least one NX application 105 (described below) on at least one server 120 to at least one honeypot 110 and at least one entity's network 125 (e.g., a private network of a company). The NX application 105 can determine if one or more assets 115 (e.g., computers) on the at least one entity's network 125 is infected with malware. It should be noted that the asset can be a simple asset (e.g., mainframe hardware, storage) or a complex asset (e.g., licensed software).

The determination of whether an asset is infected can comprise: collecting NX domain names from at least one honeypot and at least one asset; using the honeypot NX domain names to create training vectors; using the real network NX domain names to create testing vectors; classifying the testing vectors as benign vectors or malicious vectors; and classifying the at least one asset in the at least one real network as infected if the NX testing vector created from the real network NX domain names is classified as a malicious vector. (It should be noted that the testing vectors can be classified using: simple internal assets infected with known malware; simple internal assets infected with unknown malware; or complex internal network assets; or any combination thereof.)

NX domain name information is useful because some malware takes advantage of existing domain name system (DNS) services such as free domain testing (e.g., determining whether a new domain name is available). Such malware can use a domain name generator that employs a seed, such as the date, together with an algorithm to generate a set of domain names. The command and control (C&C) can try to register the generated domain names until a registrable subset of domain names has been identified. An infected computer can then use those daily-generated set of domain names in order to establish a new communication channel with the C&C. The victim computers will employ the same seed (i.e. date) and algorithm to generate the same set of domain names. The victim computers will then use the generated domain names in attempts to contact the C&C computer. Eventually, each victim computer will find a domain name that was registered for the C&C computer to enable daily communication between the C&C computer and the victim computers. By changing the domain name for the C&C computer (e.g., daily), it becomes difficult to statically black list the domain names or the IP addresses of the C&C computer(s).

Thus, malware which uses the above domain name resolution to establish communication with a C&C can produce many NX-Domains (NXs), which can be domain names that have not been registered with an authoritative DNS and can be observable at a recursive DNS server (“RDNS”). RDNS servers map domain names to IP addresses, also called “resolving DNS queries”. If such a mapping between a domain name and an IP address doesn't exist, the RNDS can send back to the initiator of the DNS query a “Non-Existence” response. The Non-Existence response can indicate that the domain name does not have an IP address, and is thus an NX-Domain (NX). Monitoring the NXs observable at a RDNS can provide the ability to collect all possible NXs generated from all computers connected to the RDNS.

FIG. 2 illustrates a method for creating training vectors, according to one embodiment. Referring to FIG. 2, in 205, malware NXs can be collected from at least one honeypot (e.g., an Internet-attached server that acts as a decoy, luring in potential hackers in order to study their activities and monitor how they are able to break into a system) by an NX application 105 and grouped into sets of, for example, 10.

The malware NXs can be collected so that a classifier can be trained in a controlled environment to recognize different categories of infected computers. For example, FIG. 5 illustrates a honeypot network configuration. In this example, the virtual machine names “kritis” operates as an internal virtual gateway for the virtual machines dns01, dns02 and dns03, which are infected with malware (e.g., sinowal worm, bobax worm). By monitoring the DNS traffic that originates from infected virtual machines dns01, dns02, and dns03, a pure seed of malware domain names can be obtained.

In FIG. 5. the computer called “minoas” can act as an open recursive DNS ((ORDNS), which can be a rDNS server willing to resolve a domain name for any host in the Internet—inside or outside its network) and as an authoritative DNS server for root DNS servers. By doing this, the minoas computer can provide NXs that appear to originate from the root DNS servers. This can force the malware to lookup the next domain name and not stop probing since the minoas computer does not allow the malware to contact the root servers. As indicated earlier, the malware needs to make contact with the C&C at least one time during a set period (e.g., daily). Thus, by providing NX answers to any domain name that the malware requests, the “minoas” computer can cause the malware to keep looking up all the generated domain names (e.g., 10,000) because no successful C&C connection will take place. In this way, all 10,000 domain names can be observed and can be used to train a statistical class that can identify malware based only on this traffic.

Referring again to FIG. 5, the “kritis” computer can be configured to give free Internet access to the dns01, dns02, and dns03 computers for one hour, and for the next eight hours to redirect the DNS traffic to the “minoas” computer. A simple IP table firewall “rotating rule” at the gateway point (e.g., at the “kritis” computer) can be used to do this.

The VMNET34 computer in FIG. 5 can be a virtual network connection between the virtual machines dns01, dns02, and dns03, and the virtual gateway “kritis”.

Referring back to FIG. 2, in 210, training vectors can be created by taking each set of, for example, 10 domain names and computing various statistical values and putting the various statistical values in a vector. Example statistics are illustrated in FIG. 6, which is described in more detail below.

Those of ordinary skill in the art will see that training vectors can be created in many other ways, in addition to collecting NXs from honeypots, as described above.

FIG. 3 illustrates a method for creating testing vectors, according to one embodiment. In 305, NXs are collected from a real network. In 310, the NXs from the real network can be used to create testing vectors by taking each set of, for example, 10 NX domain names and computing various statistical values and putting the various statistical values in a vector. (It should be noted that both the honeypot NXs and the real network NXs can be grouped in any number, and any algorithm can be used to group the sets.) It is not known if the testing NXs are malware or not. Thus, in 315, the testing vectors can be classified as benign vectors or malicious vectors by comparing testing vectors to training vectors. A classifier can use the knowledge obtained from the statistical information from the training vectors and compare it to the statistical information from the testing vectors to identify each different malware family in the testing NX vectors. FIG. 8 illustrates several types of classifiers that can be used to compare the vector information and identify different malware families. In particular, FIG. 8 illustrates the following classifiers: Naïve Bayes, LAD Tree, Multi-Layer Perception, Logistic Regression, and IBK Lazy. Those of ordinary skill in the art will see that many other types of classifiers can also be used. In addition, as explained in more detail below with respect to FIG. 8, a meta-classifier can use many different types of classifiers. In some embodiments, as also described in more detail below with respect to FIG. 8, a confidence score can also be given for each classifier, as well as for the meta-classifier.

For example, an absolute timing sequence, which can list the domain names in the order that they are received, can be used to group together an example set of ten NX domain names (e.g., from a real network):

• fpemcjfbv.com
• odkigktjzv.biz
• odkigktjzv.biz.ebay.com
• l-sjn-sevans.cal.paypal.com
• xvoal.com
• ymtaiwwprpq.biz
• ymtaiwwprpq.biz.ebay.com
• bcbkdfkg.net
• bcbkdfkg.net.ebay.com
• okxixsulas.net

An example of various statistical values that can be computed for the set of NX domain names is illustrated in FIG. 6. Note that many other types of statistical values can be computed, and that the vector can have more or less statistical values than that called for in FIG. 6 (e.g., 17). Thus, for the example of 10 NX domain names provided above, the following statistical values can be computed. It should be noted that some or all of these statistical values can be computed. In addition, other statistical values can be computed and used.

• • The average of domain name length (not including “.”) (e.g., the domain name length of the first domain name is 13). [Value ≈12.8333]
  • The standard deviation of the domain name length. [Value ≈1.9507]
  • The number of different Top Level Domains (TLDs). [Value ≈3.0]
  • The length of the longest domain name (excluding the TLD). [Value ≈24.0]
  • The median of the frequency of each unique character across the entire set of domain names (e.g., the frequency of “o” across the entire set of 10 domain names above is 10). [Value ≈2.0]
  • The average frequency of each unique character across the entire set of domain names. [Value ≈2.2083]
  • The standard deviation of the frequency of each unique character across the entire set of domain names. [Value ≈0.9565]
  • The median of the frequency of each unique 2-gram across the entire set of 10 domain names (e.g., the frequency of “fp” across the entire set of 10 domain names above is 1) (Note that if there is a “.” (e.g., “v.c”) between two characters, the frequency is counted as 0.) [Value ≈0.9565]
  • The average of the frequency of each unique 2-gram across the entire set of 10 domain names. [Value ≈1.0]
  • The standard deviation of the frequency of each unique 2-gram across the entire set of 10 domain names. [Value ≈1.0]
  • The frequency of .com TLDs over the frequency of the other of TLDs. [Value ≈1.5]
  • The median of the frequency of each unique 3-gram across the entire set of 10 domain names. [Value ≈0.3333]
  • The average of the frequency of each unique 3-gram across the entire set of 10 domain names. [Value ≈1.0]
  • The standard deviation of the frequency of each unique 3-gram across the entire set of 10 domain names. [Value ≈1.0]
  • The median count of unique TLDs (excluding .com). [Value ≈2.0]
  • The average count of unique TLDs (excluding .com). [Value ≈2.0]
  • The standard deviation for the different frequencies for each different TLD in the set of domain names. [Value ≈2.0]

The various statistical values for each set of 10 domain names from the real network NXs can be put in a vector. An example illustrating the domain names being transformed to statistical vectors, using the statistical values set forth in FIG. 6, is illustrated in FIG. 7. Referring to FIG. 7, in 705, the 10 domain names used to create the vector are listed. Note that all of these domain names can come from one particular asset 115 (e.g., an infected computer) in the real network 125:

• fpemcjfbv.com
• odkigktjzv.biz
• odkigktjzv.biz.interl.com
• l-sjn-sevans.cal.intern2.com
• xvoal.com
• ymtaiwwprpq.biz
• ymtaiwwprpq.biz.interl.com
• bcbkdfkg.net
• bcbkdfkg.net.interl.com
• okxixsulas.net

The 17 statistical values corresponding to the statistical values found in FIG. 6 are illustrated in the vector 710: [12.8333, 1.9507, 3.0, 24.0, 2.0, 2.2083, 0.9565, .9565, 1.0, 1.0, 1.5, 0.3333, 1.0, 1.0, 0.0, 2.0, 2.0, 2.0].

The NX application 105 can then utilize a meta-classifier to classify the testing vectors. The meta-classifier is a hybrid classifier and can comprise several generic classifiers. The various generic classifiers can be used (e.g., in parallel) to capture various different statistical properties which can potentially lower false positives (FP) and increase true positives (TP).

For example, FIG. 8 illustrates a meta-classifier that is comprised of five different classifiers: the Naïve Bayes classifier 805, the LAD Tree classifier 810, the Multi-Layer Perception Neural Network classifier 815, the Logistic Regression classifier 820, and the IBK Lazy Classifier 825. The maximum probability includes the classification (given by a particular classifier for the malware) and the probability of this classification being correct. Thus, for example, five different types of classifiers can be used to classify the malware as follows:

• Classifier 1 (Naive Bayes Meta.) is: notknown (Confidence: 1)
• Classifier 2 (Multi Layer Per. Meta.) is: conficker-B (Confidence: 0.985572986223)
• Classifier 3 (Logistic Regression Meta.) is: conficker-B (Confidence: 0.374297598072)
• Classifier 4 (LADtree Meta.) is: conficker-B (Confidence: 0.220571723953)
• Classifier 5 (Lazy IB1 Meta.) is: conficker-B (Confidence: 1)

The majority voting can take the many classifications and determine which classification the majority of classifiers found. Thus, for the example above, conficker-B was the classification the majority of classifiers classified the malware as. The final class is the final classification based on the majority voting, which is conficker-B.

It should be noted that the meta-classifier can use any number and any type of known or unknown classifier, including, but not limited to, the above classifiers. The Naïve Bayes classifier can use estimator classes. Numeric estimator precision values can be chosen based on analysis of the training data. The LAD tree classifier can generate a multi-class alternating decision tree using a LogitBoost strategy. The Multi-Layer Perception Neural Network classifier can use back-propagation to classify instances. The Logistic Regression classifier can build linear logistic regression models. LogitBoost with simple regression can function as a base learner and can be used for fitting the logistic models. The IBK Lazy classifier can use normalized Euclidean distance to find the training instance closest to the given test instance, and can predict the same class as the training instance. If multiple instances have the same (smallest) distance to the test instance, the first one found can be used.

Additional information about all of the above classifiers can be found in Richard O. Duda et al., PATTERN CLASSIFICATION (2nd. Edition), which is herein incorporated by reference. Further information about the IBK Lazy classifier can be found in Niels Landwehr et al, LOGISTIC MODEL TREES (2005), which is also herein incorporated by reference.

For example, each classifier in the meta-classifier can classify vector 710 as follows:

• Classifier 1 (Naive Bayes Meta.) is: notknown (Confidence: 1)
• Classifier 2 (Multi Layer Per. Meta.) is: conficker-B (Confidence: 0.985572986223)
• Classifier 3 (Logistic Regression Meta.) is: conficker-B (Confidence: 0.374297598072)
• Classifier 4 (LADtree Meta.) is: conficker-B (Confidence: 0.220571723953)
• Classifier 5 (Lazy IB1 Meta.) is: conficker-B (Confidence: 1)

Using the classification of the vector by each classifier, if a confidence threshold is set to be >=0.9 (note that this value can be set by the user), the meta-classifier can classify the vector (or statistical instance) as follows:

Instance 1 Meta classification detection result: conficker-B with majority voting value: 4 with confidence (med/std): (0.985572986223/0.345308923709). This means that a majority of four (out of five) of the classifiers found the vector to be classified as conficker-B. The median confidence score is the median of all five of the confidence scores, divided by the standard deviation of all five of the classifiers. It should be noted that, because the confidence threshold is set to be >=0.9, this number is only meaningful if the median confidence score is >=0.9.

FIG. 9 illustrates False Positive (FP) and True Positive (TP) classification results from the meta-classifier of FIG. 8 to one of six different malware classes: conficker-A, conficker-B, conficker-C, sinowal, bobax, and unknown. FIG. 9 indicates a FP value and a TP value for each type of malware. The FP rate is the False Positive detection rates for each different class. The TP rate is the True Positives detection rates for each different class. The FP rate can correspond to the percentage of vectors mistakenly classified as malicious which were actually benign. The TP rate corresponds to the percentage of vectors classified as malicious that were actually malicious. The following article, which is herein incorporated by reference, describes FP and TP rates in more detail: Axelsson. S., The Base-Rate Fallacy and the Difficulty of Intrusion Detection, ACM TRANS. INF. SYST. SECUR. 3, 3 (Aug. 2000), 186-205.

It should be noted that the meta-classifier can be independent from the manner in which the NXs are collected. It is only necessary to keep a mapping between the internal asset that the NXs originated from. The detection flow is satisfied as long as the monitoring system in the real network collects NXs from the same internal asset and groups them into sets of 10 using the absolute timing sequence. This is because the classifier can be trained to detect such behavior. Thus, the trained classifier can utilize domain names collected in the same way in real time.

FIG. 9 also illustrates how domain names from known malware (e.g., sinowal, bobax, conficker) can be classified by the meta-classifier using information learned from the training set. Domain names that do not match the training set can be classified as “notknown” or “unknownBot”. Hand verification or other methods can be used to classify the “notknown” domain names.

It should be noted that if many NXs are classified as “unknown”, either a DNS issue causes such characterization, or the NXs are from malware where little or no information about the malware is known (e.g., a new worm). DNS issues can include a DNS outage or DNS misconfiguration. If a DNS issue is the cause of the high number of “unknown” classifications, the NXs can be classified as legitimate using for example, alexa.com, or a passive DNS feed. A passive DNS feed can be a technology which constructs zone replicas without cooperation from zone administrators, based on captured name server responses (see, e.g., F. Weimer, Passive DNS Replications, http://www.enyo.de/fw/software/dnslogger/2007, which is herein incorporated by reference). An example of a passive DNS feed is a DNSParse, which can be, for example, an implementation of the passive DNS technology by the University of Auckland in New Zealand (see, e.g., https://dnsparse.insec.auckland.ac.nz/dns/2007, which is herein incorporated by reference).

FIG. 10 illustrates an example of how to identify similar patterns in NXs and use those patterns to train a new class (e.g., an unknown-bot class). For example, reviewing the NXs of FIG. 10, a person or computer program could determine malware patterns such as, but not limited to: a size of 8 (after www) with a top level domain of .com.

While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example, and not limitation. It will be apparent to persons skilled in the relevant art(s) that various changes in form and detail can be made therein without departing from the spirit and scope of the present invention. Thus, the present invention should not be limited by any of the above-described exemplary embodiments.

In addition, it should be understood that the figures described above, which highlight the functionality and advantages of the present invention, are presented for example purposes only. The architecture of the present invention is sufficiently flexible and configurable, such that it may be utilized in ways other than that shown in the figures.

Further, the purpose of the Abstract of the Disclosure is to enable the U.S. Patent and Trademark Office and the public generally, and especially the scientists, engineers and practitioners in the art who are not familiar with patent or legal terms or phraseology, to determine quickly from a cursory inspection the nature and essence of the technical disclosure of the application. The Abstract of the Disclosure is not intended to be limiting as to the scope of the present invention in any way.

Finally, it is the applicant's intent that only claims that include the express language “means for” or “step for” be interpreted under 35 U.S.C. 112, paragraph 6. Claims that do not expressly include the phrase “means for” or “step for” are not to be interpreted under 35 U.S.C. 112, paragraph 6.

Claims

1. A method of analysis, comprising: collecting, using at least one decoy virtual machine, honeypot NX domain names from at least one known infected asset in at least one real network, the honeypot NX domain names being domain names that are not registered;collecting, using the at least one decoy virtual machine, real network NX domain names from at least one asset in the at least one real network;grouping the honeypot NX domain names and the real network NX domain names based on statistical similarities;creating at least one training vector, wherein the at least one training vector is created by: computing various statistical values for at least one group of the honeypot NX domain names, andcollecting the various statistical values for the at least one group of the honeypot NX domain names in at least one vector;creating, using the real network NX domain names, a plurality of testing vectors, wherein the plurality of testing vectors are created by: computing various statistical values for at least one group of the real network NX domain names, andcollecting the various statistical values for the at least one group of the real network NX domain names in the plurality of testing vectors;classifying each of the testing vectors as benign vectors or malicious vectors based on the at least one training vector; andclassifying the at least one asset in the at least one real network as infected if at least one of the plurality of testing vectors is classified as a malicious vector.

2. The method of claim 1, further comprising classifying previously unclassified malware from the honeypot NX domain names.

3. The method of claim 1, wherein only domain name system (DNS) NX domain name information is utilized to classify the at least one asset as infected.

4. The method of claim 1, wherein only NX domain traffic is utilized.

5. The method of claim 1, wherein a meta-classifier is utilized to classify the testing vectors as benign vectors or malicious vectors.

6. The method of claim 5, wherein the meta-classifier provides intelligence for identifying new malware.

7. The method of claim 1, wherein the classifying of the testing vectors is done using at least one meta-classifier, the at least one meta-classifier comprising at least one generic classifier.

8. The method of claim 1, further comprising classifying previously classified malware from the honeypot NX domain names.

9. The method of claim 1, wherein the honeypot NX domain names collected from the at least one known infected asset and the real network NX domain names collected from the at least one asset are grouped into sets of 10 using absolute timing sequence information.

10. A system of analysis, comprising: at least one computer connected to at least one network;at least one application executing in the at least one computer, the at least one application configured for:collecting, using at least one decoy virtual machine, honeypot NX domain names from at least one known infected asset in at least one real network, the honeypot NX domain names being domain names that are not registered;collecting, using the at least one decoy virtual machine, real network NX domain names from at least one asset in the at least one real network;grouping the honeypot NX domain names and the real network NX domain names based on statistical similarities;creating at least one training vector, wherein the at least one training vector is created by: computing various statistical values for at least one group of the honeypot NX domain names, andcollecting the various statistical values for the at least one group of the honeypot NX domain names in at least one vector;creating, using the real network NX domain names, a plurality of testing vectors, wherein the plurality of testing vectors are created by: computing various statistical values for at least one group of the real network NX domain names, andcollecting the various statistical values for the at least one group of the real network NX domain names in the plurality of testing vectors;classifying each of the testing vectors as benign vectors or malicious vectors based on the at least one training vector; andclassifying the at least one asset in the at least one real network as infected if at least one of the plurality of testing vectors is classified as a malicious vector.

11. The system of claim 10, wherein the at least one application is further configured for classifying previously unclassified malware from the honeypot NX domain names.

12. The system of claim 10, wherein only domain name system (DNS) NX domain name information is utilized to classify the at least one asset as infected.

13. The system of claim 10, wherein only NX domain traffic is utilized.

14. The system of claim 10, wherein a meta-classifier is utilized to classify the testing vectors as benign vectors or malicious vectors.

15. The system of claim 14, wherein the meta-classifier provides intelligence for identifying new malware.

16. The system of claim 10, wherein the classifying of the testing vectors is done using at least one meta-classifier, the at least one meta-classifier comprising at least one generic classifier.

17. The system of claim 10, wherein the at least one application is further configured for classifying previously classified malware from the honeypot NX domain names.

18. The system of claim 10, wherein the honeypot NX domain names collected from the at least one known infected asset and the real network NX domain names collected from the at least one asset are grouped into sets of 10 using absolute timing sequence information.