Ranking cybersecurity alerts from multiple sources using machine learning

Abstract

The present disclosure relates to a machine-learning system, method, and computer program for ranking security alerts from multiple sources. The system self-learns risk levels associated with alerts by calculating risk probabilities for the alerts based on characteristics of the alerts and historical alert data. In response to receiving a security alert from one of a plurality of alert-generation sources, the alert-ranking system evaluates the security alert with respect to a plurality of feature indicators. The system creates a feature vector for the security alert based on the feature indicator values identified for the alert. The system then calculates a probability that the security alert relates to a cybersecurity risk in the computer network based on the created feature vector and historical alert data in the network. The system ranks alerts from a plurality of different sources based on the calculated cybersecurity risk probabilities.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

This invention relates generally to cybersecurity analytics in computer networks, and, more specifically, to ranking cybersecurity alerts from multiple sources using machine learning.

2. Description of the Background Art

Organizations are faced with the ever-increasing risks from security threats. Some cyberattacks are perpetrated by outsiders, while others involve insiders. Organizations typically run various cybersecurity products from different vendors. For example, one vendor may detect for malware installed on user devices, and another vendor may model and analyze user behavior to detect anomalies. Each of the different products generate alerts when a suspicious activity for which they are monitoring occurs. For a large organization with many employees, this can result in a large volume of alerts on a daily basis.

The analysts (e.g., Tier-1 analysts, Tier-3 analysts) that process these security alerts are often overwhelmed by the number of alerts. Because of the high volume of alerts, they are not able to quickly decide which alerts are not interesting and which are worthy of further investigation. A cybersecurity analyst may face over 10,000 alerts in a month and over half of them may be false positives. At many organizations, a significant percentage (e.g., 25-75%) of alerts are simply ignored because the organization cannot keep up with the alert volume. Therefore, there is demand for a system that ranks alerts from different sources so that analysts can prioritizes their attentions and focus on the alerts most likely to relate to a malicious event in the network. Such a system would greatly improve the efficiency of these analysts and enable the analysts to better monitor for cybersecurity risks.

SUMMARY OF THE DISCLOSURE

The present disclosure relates to a machine-learning system, method, and computer program for ranking cybersecurity alerts from multiple alert-generation sources in a network. The system uses past alert data to self-learn risk levels associated with alerts from different sources. Specifically, as described in more detail below, for each alert, the system calculates the probability that the alert is a cybersecurity risk based on characteristics of the alert and historical alert data from the various alert-generation sources in the network being monitored.

Multiple network-monitoring applications generate security alerts, which are received by the alert ranking system. In response to receiving a security alert from one of a plurality of alert-generation sources, the alert-ranking system evaluates the security alert with respect to a plurality of feature indicators. The system identifies values for the feature indicators with respect to the alert, and creates a feature vector representation of the alert based on the identified values. The system then calculates a probability that the security alert relates to a cybersecurity risk in the computer network based on the created feature vector and historical alert data in the network. In certain embodiments, the calculated risk probability is a Bayes probability calculated as a function of the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity.

A risk probability is calculated for each alert received, and the system ranks the security alerts based on the calculated risk probabilities. The ranked list includes alerts from a plurality of different network-monitoring applications, therefore providing cybersecurity analysts with a unified alert ranking system.

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 is a flowchart that illustrates a method, according to one embodiment, for ranking computer network security alerts from multiple alert-generation sources.

FIG. 2 is a table that illustrates example feature indicators for security alerts.

FIG. 3 is a table that illustrates an example ranking of security alerts.

FIG. 4 is a block diagram that illustrates an example software architecture for a cybersecurity alert-ranking system.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The present disclosure describes a machine-learning system, method, and computer program for ranking security alerts. The method is performed by a computer system that receives security alerts for a computer network monitored for cybersecurity risks and attacks (“the system”). The system learns from past data to rank alerts. As described in more detail below, the system creates a feature vector representation of each incoming alert, and then calculates the probability that malicious activity has occurred given the feature vector and past alert data. The calculated probability is used to rank the alert.

FIG. 1 illustrates a method for ranking computer network security alerts from multiple sources. The illustrated method is described in more detail below.

1. Evaluating a Security Alert with Respect to Feature Indicators

The system receives a security alert for a computer network from one of a plurality of alert-generation sources (step 110). Alerts are generated by various sources within the monitored computer network. Example sources are third-party security product vendors that produce data loss prevention alerts, web traffic alerts, and endpoint malware alerts, etc. In other cases, an alert may be generated by a user behavior analytics (UBA) or a user and entity behavior analytics (UEBA) system. An example of a UBA/UEBA cybersecurity monitoring system is described in U.S. Pat. No. 9,798,883 issued on Oct. 24, 2017 and titled “System, Method, and Computer Program for Detecting and Assessing Security Risks in a Network,” the contents of which are incorporated by reference herein.

The system evaluates the security alert with respect to multiple feature indicators and identifies values for the feature indicators with respect to the alert (step 120). The feature indicators represent features of an alert or context information for an alert. For some feature indicators, the system may determine whether the feature indictor evaluates to true or false with respect to the alert. In such cases, the value of the feature indicator with respect to the alert may be a Boolean data type corresponding to true or false. For other feature indicators, the value may be a numerical value within a range (e.g., an number representing a severity level) or a text string (e.g., the alert type or name). The table in FIG. 2 lists nine example feature indicators and the corresponding values for which the feature indicators are evaluated, but these are just examples and invention is not limited to these feature indicators.

2. Creating a Feature Vector Representation of the Alert

The system creates a feature vector representation of the alert (step 130). Specifically, the system creates a feature vector for the alert that includes the evaluated values for feature indicators with respect to the alert. Let {right arrow over (f)}_a=(f₁, f₂, . . . f_i) denote a feature vector for an alert, where there are I feature indicators and f_i is the value for the ith feature indicator.

3. Calculating a Probability that the Security Alert Relates to a Cybersecurity Risk.

The system calculates a probability that that security alert relates to a cybersecurity risk (i.e., to malicious activity in the network) based on the alert's feature vector and the historical alert data for the monitored network (step 140) In one embodiment, the risk probability is a Bayes probability calculated as a function of the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate activity. Specifically, the probability may be calculated by the system as follows.

$P\left(M❘\overrightarrow{f_a}\right)=\frac{P\left(\overrightarrow{f_a}❘M\right)P\left(M\right)}{P\left(\overrightarrow{f_a}❘M\right)P\left(M\right)+P\left(\overrightarrow{f_a}❘L\right)P\left(L\right)}$

• • Where:
  • P(M|{right arrow over (f)}_a) is the posterior probability that the alert relates to a cybersecurity risk given the feature vector {right arrow over (f)}_a;
  • P({right arrow over (f)}_a|L) is the likelihood of observing the feature vector in connection with legitimate or low-interest activity. The system learns P({right arrow over (f)}_a|L) based on the features observed in past alerts. In one embodiment, P({right arrow over (f)}_a|L) is a function of how often {right arrow over (f)}_a has been observed in the monitored network and is the calculated by the system as follows:

$P\left(\overrightarrow{f_a}❘L\right)=\frac{\mathrm{count}\mathrm{of}\mathrm{observed}\mathrm{alerts}\mathrm{each}\mathrm{with}\stackrel{\rightharpoonup }{f_a}}{\mathrm{count}\mathrm{of}\mathrm{all}\mathrm{observed}\mathrm{alerts}}$

• • P({right arrow over (f)}_a|M) is the likelihood of observing the feature vector in connection with a cybersecurity risk (i.e., in connection with malicious activity in the monitored network). In one embodiment, P({right arrow over (f)}_a|M) is assumed to be uniform in which case can be calculated by the system as follows:

$P\left(\overrightarrow{f_a}❘M\right)=\frac{1}{\mathrm{number}\mathrm{of}\mathrm{distinct}\stackrel{\rightharpoonup }{f_a}}$

• • P(M) is the a-priori probability of a cybersecurity risk. P(L) is the a-priori probability of a legitimate or low-interest activity. In one embodiment, P(M) and P(L) are set to the value 0.5.

Computationally it may not be practical for the system to calculate P({right arrow over (f)}_a|L) and P({right arrow over (f)}_a|M) based on joint modeling of all the feature indicators. For example, in the case of the feature indictors in the table in FIG. 2, “fOrg,” “fDev,” “nNewDev7d,” “nVendor7d” are binary features, each with a true/false values. Therefore, the total number of feature values is small, or at most 2⁴ possible scenarios. However, the feature indicators “alert_type” and “alert_name” may have 10 or 20 different values. If the features are modeled jointly, there are 2⁴*10*10 possible scenarios (assuming 10 different values for each), which requires a significant amount of data volume to calculate P({right arrow over (f)}_a|L) and ({right arrow over (f)}_a|M).

It is more computationally efficient to divide f_a into two or more subsets, calculate P({right arrow over (f)}_a|L) and P({right arrow over (f)}_a|M) for each subset, and then calculate the product of the probabilities for each subset. This can be denoted mathematically as follows:

$\begin{array}{c}P\left(\overrightarrow{f_a}❘L\right)=\prod _{j=1}^{J}P\left({\overrightarrow{g}}_i❘L\right)\\ P\left(\overrightarrow{f_a}❘M\right)=\prod _{j=1}^{J}P\left({\overrightarrow{g}}_i❘M\right))\end{array}$

Where there are J conditional independent groups {right arrow over (g)}_i, each consisting of a non-overlapping subset of features f_a, and where P({right arrow over (g)}_i|L) and P({right arrow over (g)}_i|M) are calculated by the system as follows:

$\begin{array}{c}P({\overrightarrow{g}}_i❘L)=\frac{\mathrm{count}\mathrm{of}\mathrm{observed}\mathrm{alerts}\mathrm{each}\mathrm{with}{\overrightarrow{g}}_i}{\mathrm{count}\mathrm{of}\mathrm{all}\mathrm{observed}\mathrm{alerts}}\\ P\left({\overrightarrow{g}}_i❘M\right)=\frac{1}{\mathrm{number}\mathrm{of}\mathrm{distinct}{\overrightarrow{g}}_i}\end{array}$

The groups may be divided based on which feature indicators are considered probabilistically independent of each other, such that each group is probabilistically independent of the other groups with respect to P({right arrow over (g)}_i|L) and P({right arrow over (g)}_i|M). For example, for the feature indicators listed in FIG. 2, the features 1-5 may be a first group, features 6-7 may be a second group, feature 8 may be a third singleton group, and feature 9 may be a fourth singleton group.

In certain embodiments, instead of assuming P({right arrow over (g)}_i|M) is uniform, human knowledge can be injected to influence the risk rankings. For example, if {right arrow over (g)}={alert_type} and we know a priori that alerts with alert_type=“web attack” are not a good malicious indicator, then an administrator of the system can set P(alert_type=‘web attack’|M) to a smaller number than the rest of the P(alert_Type|M) likelihoods.

In one embodiment, the system profiles alert data in a monitored computer network for a number of days (e.g., 7 days) prior to calculating the risk probabilities and ranking security alerts. Since the probabilities are based in part on alert history in the monitored network, this helps optimize the probability calculations.

4. Ranking Security Alerts Based on the Calculated Risk Probabilities

In one embodiment, the above-described steps are performed for each security alert received by the system to calculate a risk probability for each alert (step 150). The system ranks the security alerts based on the risk probabilities and displays the ranked alerts (steps 160-170). The alerts are preferably ranked and displayed in descending order of risk probability (i.e., highest-risk alerts are ranked highest and displayed at the top of the list). The ranked list may be limited to alerts received within a certain time period (e.g., a 24 hour window).

FIG. 3 illustrates an example of the displayed ranking in which the feature indicators are the same as those listed in the table in FIG. 2. The display includes the feature indicator values for the ranked alerts. Displaying the feature indicator values for the alert enables an analyst to easily see the features associated with high-ranking alerts. For example, the highest-ranked alert in FIG. 3 has alert_name “Hacktool.Jsprat” with alert_type “Security risk found” and the alert has the following features:

• • the first-time-seen-in-org is True (i.e., First_Org is true);
  • first-time-seen-in-device is True (i.e., First_Source is true);
  • the feature for having more than its source device+vendor tuple sees>=2 alert_types returns True (i.e., nAlertType7d is true); and
  • the alert_type is ‘Security risk found”.

In certain embodiments, the system determines the risk probability for each alert in substantially real time as alerts are received by the system, and updates the displayed ranking in substantially real time as new alerts are received by the system.

5. Example Software Architecture

FIG. 4 illustrates an example software architecture for a system 400 that performs the methods described herein. However, the methods described herein may be implemented in other systems and are not limited to system 400. An Alert Receiver module 410 receives (or retrieves) alerts generated by a plurality of different alert generation sources 405a-d. The Alert Feature Evaluator model 420 evaluates alerts with respect to a set of feature indicators and determines feature indicator values for each alert. Feature Vector Creation module 430 creates a feature vector for each alert using the feature indicator values for the alert. The Risk Probability Calculator module 440 calculates a probability that an alert relates to a cybersecurity risk based on the alert's feature vector and historical alert data. The Alert Ranking module 450 ranks alerts based on the risk probabilities calculated for the alerts. The UI Generator module 460 generates a user interface that displays alert rankings, such as the top ranked alerts.

6. General

The methods described herein 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.

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.

Claims

1. A method, performed by a computer system, for ranking computer network security alerts from multiple sources, the method comprising: (a) receiving a security alert from one of a plurality of alert-generation sources in a computer network;(b) evaluating the security alert with respect to a plurality of feature indicators to obtain feature indicator values for the security alert;(c) creating a feature vector for the security alert that includes the feature indicator values for the security alert;(d) calculating a probability that the security alert relates to a cybersecurity risk in the computer network based on the created feature vector and historical alert data in the network, wherein the probability is a Bayes probability calculated as a function of the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity, wherein calculating the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity comprises: dividing the feature vector for the alert into a plurality of non-overlapping subsets to create a plurality of subset feature vectors, for each subset feature vector, calculating a probability of seeing the subset feature vector with respect to a cybersecurity risk and a probability of seeing the subset feature vector with respect to legitimate or low-interest activity, and calculating the product of the probabilities calculated for the subset feature vectors to obtain the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity;(e) performing steps (a)-(d) for a plurality of security alerts from the plurality of alert-generation sources;(f) ranking the security alerts based on the calculated probabilities; and(g) displaying the ranked security alerts, wherein the alert ranking includes alerts from a plurality of alert-generation sources.

2. The method of claim 1, wherein displaying the alert ranking includes displaying the values of the feature indicators for the ranked alerts.

3. The method of claim 1, wherein the risk probability for alerts is calculated in substantially real time as alerts are received by the system, and wherein the alert ranking is updated in substantially real time as new alerts are received by the system.

4. 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 ranking computer network security alerts from multiple sources, the method comprising: (a) receiving a security alert from one of a plurality of alert-generation sources in a computer network;(b) evaluating the security alert with respect to a plurality of feature indicators to obtain feature indicator values for the security alert;(c) creating a feature vector for the security alert that includes the feature indicator values for the security alert;(d) calculating a probability that the security alert relates to a cybersecurity risk in the computer network based on the created feature vector and historical alert data in the network, wherein the probability is a Bayes probability calculated as a function of the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity, wherein calculating the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity comprises: dividing the feature vector for the alert into a plurality of non-overlapping subsets to create a plurality of subset feature vectors, for each subset feature vector, calculating a probability of seeing the subset feature vector with respect to a cybersecurity risk and a probability of seeing the subset feature vector with respect to legitimate or low-interest activity, and calculating the product of the probabilities calculated for the subset feature vectors to obtain the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity;(e) performing steps (a)-(d) for a plurality of security alerts from the plurality of alert-generation sources;(f) ranking the security alerts based on the calculated probabilities; and(g) displaying the ranked security alerts, wherein the alert ranking includes alerts from a plurality of alert-generation sources.

5. The non-transitory computer-readable medium of claim 4, wherein displaying the alert ranking includes displaying the values of the feature indicators for the ranked alerts.

6. The non-transitory computer-readable medium of claim 4, wherein the risk probability for alerts is calculated in substantially real time as alerts are received by the system, and wherein the alert ranking is updated in substantially real time as new alerts are received by the system.

7. A computer system for ranking computer network security alerts from multiple sources, 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: (a) receiving a security alert from one of a plurality of alert-generation sources in a computer network;(b) evaluating the security alert with respect to a plurality of feature indicators to obtain feature indicator values for the security alert;(c) creating a feature vector for the security alert that includes the feature indicator values for the security alert;(d) calculating a probability that the security alert relates to a cybersecurity risk in the computer network based on the created feature vector and historical alert data in the network, wherein the probability is a Bayes probability calculated as a function of the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity, wherein calculating the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity comprises: dividing the feature vector for the alert into a plurality of non-overlapping subsets to create a plurality of subset feature vectors, for each subset feature vector, calculating a probability of seeing the subset feature vector with respect to a cybersecurity risk and a probability of seeing the subset feature vector with respect to legitimate or low-interest activity, and calculating the product of the probabilities calculated for the subset feature vectors to obtain the probability of seeing the feature vector with respect to a cybersecurity risk and the probability of seeing the feature vector with respect to legitimate or low-interest activity;(e) performing steps (a)-(d) for a plurality of security alerts from the plurality of alert-generation sources;(f) ranking the security alerts based on the calculated probabilities; and(g) displaying the ranked security alerts, wherein the alert ranking includes alerts from a plurality of alert-generation sources.

8. The system of claim 7, wherein displaying the alert ranking includes displaying the values of the feature indicators for the ranked alerts.

9. The system of claim 7, wherein the risk probability for alerts is calculated in substantially real time as alerts are received by the system, and wherein the alert ranking is updated in substantially real time as new alerts are received by the system.