System, method, and computer program for automatically classifying user accounts in a computer network using keys from an identity management system

Abstract

The present disclosure describes a system, method, and computer program for automatically classifying user accounts within an entity's computer network, using machine-based-learning modeling and keys from an identity management system. A system uses supervised machine learning to create a statistical model that maps individual keys or sets of keys to a probability of being associated with a first type of user account (e.g., a service account). To classify an unclassified user account, the system identifies identity management keys associated with the unclassified user account. The system creates an N-dimensional vector from the keys (where N=the number of keys), and uses the vector and the statistical model to calculate a probability that the unclassified user account is the first type of user account. In response to the probability exceeding a first threshold, the system classifies the unclassified user account as the first type of user account.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

This invention relates generally to security analytics in computer networks, and, more specifically, to classifying user accounts as human accounts or service accounts based on keys from an identity management system.

2. Description of the Background Art

For user behavior modeling in IT network security analytics, it is critical to leverage contextual information to improve alert accuracy. For example, contextual information can be used to construct and evaluate context-specific rules. Some contextual information is factual, and some is derived statistically. An example of factual contextual information is the geolocation from which a current VPN event comes. An example of statistically-derived contextual information is a conclusion that an asset (e.g., a laptop or phone) is likely to belong to an executive based on historical data.

Whether an account is a human user account or a service account is useful contextual information in network security analytics. For example, if during a login session, an account is behaving as a service account, but it is known that it is a human user account, the login session may be a good candidate for an alert.

An identity management system (e.g., Open LDAP, Active Directory) maintains a directory of all accounts on an enterprise network. Each account is described by a collection of key-value pairs. “Keys” are akin to fields, but are dynamic in that they some can be specified by the enterprise. The types of keys used to describe an account are not always consistent across departments and certainly not across enterprises.

Currently, classifying an account as a human user account or a service account is done manually and requires significant human effort. An analyst reads the organization unit key from an identity management system and decides whether the key value pertains to a service account. This environment-specific effort is laborious and at best finds a subset of service accounts, leaving potentially other service accounts undiscovered. Furthermore, the process needs to be repeated as new accounts are added to the network. It would be desirable to leverage the manually-found service accounts to construct an automated classifier to probabilistically infer, using textual readout of keys from an identity management system, the status of new accounts or existing, unclassified accounts.

SUMMARY OF THE DISCLOSURE

The present disclosure describes a system, method, and computer program for automatically classifying user accounts within an entity's computer network, using machine-based-learning modeling and keys from an identity management system. The method is performed by a computer system (the “system”).

Using machine-learning-based modeling, the system creates a statistical model that maps individual keys or sets of keys to a probability of being associated with a first type of user account. The model is trained using a set of inputs and a target variable. The inputs are keys from an identity management data structure associated with user accounts manually classified as the first type of user account or a second type of user account, and the target variable is whether the user account is the first type of user account.

Once the statistical model is created, the system uses the model to automatically determining whether an unclassified user account is the first type of user account. To classify an unclassified user account, the system identifies identity management keys associated with the unclassified user account. The system then creates an N-dimensional vector of the keys, wherein N is the number of the keys associated with the unclassified user account.

The system inputs the N-dimensional vector into the statistical model to calculate a probability that the unclassified user account is the first type of user account. In response to the probability exceeding a first threshold, the system classifies the unclassified user account as the first type of user account.

In certain embodiments, there is one threshold (i.e., the first threshold). If the probability is below the first threshold, the account is classified as the second type of account.

In certain embodiments, there are two thresholds. If the probability is below a lower, second threshold, the account is classified as the second type of account. If the probability is between the first and second thresholds, the system concludes that the classification of the user account is undetermined.

In certain embodiments, the first type of user account is a service user account and the second type of user account is a human user account. In certain embodiments, the first type of user account is a human user account and the second type of account is a service user account.

In certain embodiments, the data model is constructed using Bernoulli Naïve Bayes modeling.

In certain embodiments, the keys for the unclassified user account are identified by parsing an output text file from the identity management system that corresponds to the unclassified user account.

In certain embodiments, the system performs the automated classification steps on each of the manually-classified user accounts used to train the statistical model in order to identify any mismatches between the automated classifications and the manual classifications.

BRIEF DESCRIPTION OF THE DRAWINGS

FIGS. 1A-1B are flowcharts that illustrate a method, according to one embodiment, for automatically classifying user accounts as human user accounts or service accounts.

FIGS. 2 and 3 are graphs that illustrate example classification results.

FIG. 4 is a flowchart that illustrates a method, according to one embodiment, for calculating a probability threshold.

FIG. 5 is a flowchart that illustrates a method, according to one embodiment, for lowering the Equal Error Rate of the classification results.

FIG. 6 is a table that illustrates example key-to-probability mappings in a Naïve Bayes statistical model.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

FIGS. 1A-1B illustrates a method for automatically classifying user accounts as human user accounts or service accounts within an entity's network (e.g., an enterprise's IT network). The method uses machine-learning-based modeling and keys from an identity management system to classify accounts. The method is performed by a computer system (the “system”). An entity may be an enterprise, a corporation, a government agency, or any other type of organization.

Referring to FIG. 1, the system creates a statistical model that maps individual keys, or sets of keys, from an identity management system to a probability of being associated with a service account in the entity's IT network (step 110).

The model is created using a supervised learning algorithm. The preferred algorithm is the Bernoulli Naïve Bayes algorithm, but other supervised learning algorithms, such as logistical regression algorithm, could be used. The model is trained using a set of inputs and a target variable. The inputs used to train the statistical model are identity management keys associated with manually-classified accounts. In the embodiment illustrated in FIG. 1, the target variable for the model is whether an account is a service account. In other words, the model maps identity management keys to a probability that an account is a service account. However, those skilled in the art will appreciate that the target variable could be whether the account is a human user account.

The system obtains the keys for the training data from an identity management system. Examples of identity management systems are MICROSOFT ACTIVE DIRECTORY and OPEN LDAP. In one embodiment, the system parses an LDIF text file from identity management system to find the keys associated with training accounts. A human then provides a list of accounts known as service accounts based on the knowledge of enterprise's account naming convention or IT records. Alternatively, a human may review the keys to manually classify service accounts from the keys. For example, an administrator may review the keys for a set of accounts to identify the accounts with a key or set of keys known to be specific only to service accounts at an entity. The identified accounts within the training set are manually classified as service accounts, and the other accounts within the set are manually classified, by default, as human user accounts. Whether via known service account listing or manual review effort, the administrator likely will find only a subset of the service accounts this way, but, if the identified set of service accounts is large enough, the model will be sufficiently reliable. Furthermore, as described below, an iterative process may be used to improve the reliability of the model. Using a supervised learning algorithm, the system leverages the manually classified accounts to “learn” and build the statistical model.

The table in FIG. 6 illustrates example key-to-probability mappings in a Naïve Bayes statistical model created as described above. Specifically, the table maps example individual keys to a probability that the key is associated with a service account. The keys are examples of keys associated with one entity, and keys may vary across entities. In “P(C|word)” in FIG. 6, C=Service Account.

Once the model is created, it can be used to automatically determine whether an unclassified account is a service account or human user account. It also can be used to reclassify, in an automated manner, the manually-classified training data. Referring to FIG. 1, the system obtains from, an identity management system, the identity management keys associated with the unclassified account (step 120). In one embodiment, the system parses an LDIF text file from the identity management system to obtain the textual keys. The system creates an N-dimensional vector of the keys, wherein N is the number of keys associated with the unclassified account (step 130). The system inputs the N-dimensional vector into the statistical model to calculate a probability that the unclassified user account is a service account (step 140).

FIGS. 2 and 3 are graphs that illustrate example classification results. In each graph, the x-axis 202, 302 represents the probability scores (in logarithmic form (base 10)), and the y-axis 204a, 204b, 304a, 304b represents the number of accounts (204a and 304a are the y-axis for human user accounts, and 204b and 304b are the y-axis for service accounts). The bars represent the automated classification results. In these examples, the accounts classified are the original training accounts, and the fill in the bar represents whether the bar was manually classified as a service account or a human user account (hash mark=manually classified as human user account).

Referring again to FIG. 1, if the probability calculated in step 140 is above a threshold probability, the system classifies the account as a service account (step 145, 150). A threshold can be chosen based on desired tradeoff between false positive rate and false negative rate of the classifier. An example threshold is the probability where the false positive rate equals to the false negative rate. In one embodiment, if the probability is below the threshold, the account is classified as a human user account (step 160). This embodiment is illustrated in FIG. 2 in which accounts with scores above probability threshold 210 are classified as service accounts, and accounts with scores below the threshold 210 are classified as human user accounts.

In an alternate embodiment, there are two probability thresholds, an upper threshold and a lower threshold, as illustrated in FIG. 3. Accounts with probabilities above an upper threshold 310 are classified as service accounts, and accounts with probabilities below a lower threshold 305 are classified as human user accounts (step 160). The system concludes that accounts with probabilities between the two thresholds 305, 310 are undetermined. Line 315 represents the location of the EER rate, which is discussed below.

In FIGS. 1-3, the positive class is the service account in that the model predicts the probability that an account is a service account. However, those skilled in the art will appreciate, that the system could be easily configured such that the positive class is a human user account. In such case, the model is trained to predict the probability that an account is a human user account. Moreover, the above method can be applied to classify other types of accounts (e.g., a first type of account and a second type of account) using identity management keys.

In the embodiment described with respect to FIG. 1, the system classifies accounts using just the keys from an identity management system, not the key-value pairs. However, in alternate embodiment, key-value pairs are used to classify accounts. In such case, the statistical model is trained using key-value pairs and predicts the likelihood each key-value pair, or a set of key-value pairs, is a certain type of account.

FIG. 4 illustrates a method, according to one embodiment, for calculating a probability threshold, such as threshold 210 in FIG. 2. The system performs steps 120-140 on the training data (i.e., the manually-classified accounts used to train the model) to automatically classify the training data accounts (step 410). These counts were previously classified manually to train the model, but in step 410 the system uses the model to automatically reclassify the accounts. The system identifies the probability score at which the false positive (FP) rate is equal to the false negative (FN) rate (step 420). This is known as the Equal Error Rate (EER). The FP rate is the percentage of negative accounts misclassified by the system as positives, assuming that the manual classifications were correct. The FN rate is the percent of positive accounts misclassified by the system as negatives, assuming that the manual classifications were correct.

If the positive scenario is service accounts (i.e., the system predicts the probability that an account is a service account), then the FP rate is the percentage of human user accounts misclassified as service accounts by the system, assuming that the manual classifications were correct. Likewise, the FN rate is the percentage of service accounts misclassified by the system as human user accounts, assuming the manual classifications were correct. If only one threshold will be used by the system to classify accounts (e.g., threshold 210), the threshold probability is the EER probability (step 430). If two thresholds are used (e.g., thresholds 305, 310), the lower threshold is set a certain amount below the EER score, and the upper threshold is set a certain amount above the EER score (step 430). In such case, the lower threshold gives x amount of the FN rate, and the upper threshold gives y amount of the FP rate, wherein x and y are separately controllable based on the EER. For example, assume that, at the EER, the FP and FN rates are 12%. The upper and lower thresholds are set to probability scores on the x axis respectively associated with lower FP and FN rates. The lower threshold might be set to an FN rate of 6% and the upper threshold might be set to a FP rate of 6%. The percentages need not be equal.

An iterative classification process may be used to lower the EER rate and increase the accuracy of the statistical model. FIG. 5 illustrates an example of such a process. The system identifies the accounts that were manually-classified as the negative case (e.g., human user accounts), but that the statistical model predicts are the positive case (e.g., service accounts) (step 510). In FIG. 2, these would be the accounts represented by the bars with hash marks 240 that are to the right of threshold 210 (i.e., the accounts that were manually classified as human user accounts, but predicted to be service accounts by the system). An analyst reviews the identified accounts again, and updates the manual classifications as applicable. The system obtains the updated manual classifications and retrains the statistical model (steps 520, 530). The system then inputs the updated training data into the retrained statistical model and reclassifies the training accounts accordingly (step 540). This results in a lower EER rate and higher classification accuracy, while minimizing the amount of accounts that need to be manually reviewed (step 550). The probability threshold(s) are adjusted in accordance with the probability score of the new EER. This process may be repeated iteratively.

The methods described with respect to FIGS. 1-5 are embodied in software and performed by a computer system (comprising one or more computing devices) executing the software. A person skilled in the art would understand that a computer system has one or more memory units, disks, or other physical, computer-readable storage media for storing software instructions, as well as one or more processors for executing the software instructions.

The account classification results may be used in context-specific rules in security analytics for computer networks. For example, an alert may be raised if an account classified by the methods herein as a human user account is behaving like a service account.

As will be understood by those familiar with the art, the invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. Accordingly, the above disclosure is intended to be illustrative, but not limiting, of the scope of the invention, which is set forth in the following claims.

Claims

1. A method, performed by a computer system, for automatically classifying user accounts in an entity's IT network, wherein the user accounts are classified using identity management key-value pairs from an identity management data structure, the method comprising: training a statistical model to map individual identity management key-value pairs or sets of identity management key-value pairs to a probability of being associated with a service user account, wherein a key in the identity management key-value pair is a textual string that represents a field in a directory, maintained by an identity management system, comprising one or more accounts on the entity's IT network, wherein a value in the identity management key-value pair is a corresponding value to the field in the directory, and wherein the statistical model is trained using a set of inputs and a target variable and wherein training the model comprises: parsing account data from an output text file stored in or hosted on the identity management system associated with user accounts manually classified as the service user accounts or human user accounts to obtain dynamically-specified identity management key-value pairs that are used as the inputs in the statistical model, andsetting the target variable in the statistical model to be whether the user account is a service user account;using machine-learning-based modeling to automatically determine whether an unclassified user account is a service user account by performing the following: identifying identity management key-value pairs, from the identity management system, associated with the unclassified user account,representing the unclassified user account as an N-dimensional vector of the identity management key-value pairs, wherein N is the number of the identity management key-value pairs associated with the unclassified user account,inputting the N-dimensional vector into the statistical model to calculate a probability that the unclassified user account is a service user account, andin response to the probability exceeding a threshold, classifying the unclassified user account as a service user account;using account classification results from the machine-learning-based modeling to construct and evaluate context-specific rules, wherein the context-specific rules identify one or more user accounts that are classified as service user account(s) but are known in the system to be human user account(s), wherein for the one or more user accounts that are classified as service user account(s) but are known in the system to be human user account(s), performing the following steps: identifying a probability score associated with an equal error rate (EER), wherein the EER is the rate at which false positives equal false negatives,setting the threshold to the probability score associated with the EER, andin response to the probability exceeding the threshold, classifying the human user account(s) as service user account(s); andusing the context-specific rules to improve security analytics alert accuracy in an IT network.

2. The method of claim 1, further comprising: in response to the probability being below the threshold, classifying the unclassified user account as a human user account.

3. The method of claim 1, further comprising: in response to the probability being below a second, lower threshold, classifying the unclassified user account as a human user account; andin response to the probability being between the first and second thresholds, concluding that the classification of the user account is undetermined.

4. The method of claim 1, wherein the statistical model is constructed using Bernoulli Naïve Bayes modeling.

5. The method of claim 1, further comprising: performing the automated classification steps on each of the manually-classified user accounts used to train the statistical model in order to identify any mismatches between the automated classifications and the manual classifications.

6. A non-transitory computer-readable medium comprising a computer program, that, when executed by a computer system, enables the computer system to perform the following method for automatically classifying user accounts in an entity's IT network, wherein the user accounts are classified using identity management key-value pairs from an identity management data structure, the method comprising: training a statistical model to map individual identity management key-value pairs or sets of identity management key-value pairs to a probability of being associated with a service user account, wherein a key in the identity management key-value pair is a textual string that represents a field in a directory, maintained by an identity management system, comprising one or more accounts on the entity's IT network, wherein a value in the identity management key-value pair is a corresponding value to the field in the directory, and wherein the statistical model is trained using a set of inputs and a target variable and wherein training the model comprises: parsing account data from an output text file stored in or hosted on the identity management system associated with user accounts manually classified as the service user accounts or human user accounts to obtain dynamically-specified identity management key-value pairs that are used as the inputs in the statistical model, andsetting the target variable in the statistical model to be whether the user account is a service user account;using machine-learning-based modeling to automatically determine whether an unclassified user account is a service user account by performing the following: identifying identity management key-value pairs, from the identity management system, associated with the unclassified user account,representing the unclassified user account as an N-dimensional vector of the identity management key-value pairs, wherein N is the number of the identity management key-value pairs associated with the unclassified user account,inputting the N-dimensional vector into the statistical model to calculate a probability that the unclassified user account is a service user account, andin response to the probability exceeding a threshold, classifying the unclassified user account as a service user account;using account classification results from the machine-learning-based modeling to construct and evaluate context-specific rules, wherein the context-specific rules identify one or more user accounts that are classified as service user account(s) but are known in the system to be human user account(s), wherein for the one or more user accounts that are classified as service user account(s) but are known in the system to be human user account(s), performing the following steps: identifying a probability score associated with an equal error rate (EER), wherein the EER is the rate at which false positives equal false negatives,setting the threshold to the probability score associated with the EER, andin response to the probability exceeding the threshold, classifying the human user account(s) as service user account(s); andusing the context-specific rules to improve security analytics alert accuracy in an IT network.

7. The non-transitory computer-readable medium of claim 6, further comprising: in response to the probability being below the threshold, classifying the unclassified user account as a human user account.

8. The non-transitory computer-readable medium of claim 6, further comprising: in response to the probability being below a second, lower threshold, classifying the unclassified user account as a human user account; andin response to the probability being between the first and second thresholds, concluding that the classification of the user account is undetermined.

9. The non-transitory computer-readable medium of claim 6, wherein the statistical model is constructed using Bernoulli Naïve Bayes modeling.

10. The non-transitory computer-readable medium of claim 6, further comprising: performing the automated classification steps on each of the manually-classified user accounts used to train the statistical model in order to identify any mismatches between the automated classifications and the manual classifications.

11. A computer system for automatically classifying user accounts in an entity's IT network, wherein the user accounts are classified using identity management key-value pairs from an identity management data structure, the system comprising: one or more processors;one or more memory units coupled to the one or more processors, wherein the one or more memory units store instructions that, when executed by the one or more processors, cause the system to perform the operations of: training a statistical model to map individual identity management key-value pairs or sets of identity management key-value pairs to a probability of being associated with a service user account, wherein a key in the identity management key-value pair is a textual string that represents a field in a directory, maintained by an identity management system, comprising one or more accounts on the entity's IT network, wherein a value in the identity management key-value pair is a corresponding value to the field in the directory, and wherein the statistical model is trained using a set of inputs and a target variable and wherein training the model comprises: parsing account data from an output text file stored in or hosted on the identity management system associated with user accounts manually classified as the service user accounts or human user accounts to obtain dynamically-specified identity management key-value pairs that are used as the inputs in the statistical model, andsetting the target variable in the statistical model to be whether the user account is a service user account;using machine-learning-based modeling to automatically determine whether an unclassified user account is a service user account by performing the following: identifying identity management key-value pairs, from the identity management system, associated with the unclassified user account,representing the unclassified user account as an N-dimensional vector of the identity management key-value pairs, wherein N is the number of the identity management key-value pairs associated with the unclassified user account,inputting the N-dimensional vector into the statistical model to calculate a probability that the unclassified user account is a service user account, andin response to the probability exceeding a threshold, classifying the unclassified user account as a service user account;using account classification results from the machine-learning-based modeling to construct and evaluate context-specific rules, wherein the context-specific rules identify one or more user accounts that are classified as service user account(s) but are known in the system to be human user account(s), wherein for the one or more user accounts that are classified as service user account(s) but are known in the system to be human user account(s), performing the following steps: identifying a probability score associated with an equal error rate (EER), wherein the EER is the rate at which false positives equal false negatives,setting the threshold to the probability score associated with the EER, andin response to the probability exceeding the threshold, classifying the human user account(s) as service user account(s); andusing the context-specific rules to improve security analytics alert accuracy in an IT network.

12. The system of claim 11, further comprising: in response to the probability being below the threshold, classifying the unclassified user account as a human user account.

13. The system of claim 11, further comprising: in response to the probability being below a second, lower threshold, classifying the unclassified user account as a human user account; andin response to the probability being between the first and second thresholds, concluding that the classification of the user account is undetermined.

14. The system of claim 11, wherein the statistical model is constructed using Bernoulli Naïve Bayes modeling.

15. The system of claim 11, further comprising: performing the automated classification steps on each of the manually-classified user accounts used to train the statistical model in order to identify any mismatches between the automated classifications and the manual classifications.