Detecting Brute Force Attacks
By using a hybrid model to identify the distribution of multimodal operation parameter values in the network, combining the source Internet protocol address and username information, high accuracy detection and defense against brute force attacks is achieved, and the difficulty of brute force attack detection in the existing technology is solved. Big problem, lowering false positive rates and improving network security.
Patent Information
- Application Number
- CN202080086147.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-03-27
- Filing Date
- 2020-12-11
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2040-12-11
AI Technical Summary
The existing technology is difficult to effectively detect and defend against brute force attacks, especially in the absence of complex vulnerabilities. Attackers can brute force attacks through simple tools, leveraging weak passwords and servers with lack of multi-factor authentication for ransomware and other malicious activities. .
By modeling the distribution of operational parameter values, the multimodality of the computer network is identified using a hybrid model, and detection is adjusted according to dynamic thresholds to reduce false positives while providing accurate detection of brute-force attacks. This method combines source Internet protocol address and username information to generate alerts and programmatically reconfigure network security devices to mitigate attacks.
It improves the accuracy and accuracy of brute-force attack detection, reduces the false alarm rate, and can automatically adjust network security policies to prevent attacks and effectively prevent network intrusion.
Smart Images

Figure CN114788228B_ABST
Abstract
Description
Background Art
[0001] Computers exposed to the Internet are attractive targets for attackers because they present a simple and effective way to gain access to a network. Performing a brute force attack does not necessarily require a high level of expertise or sophisticated exploitation of vulnerabilities. Instead, in some scenarios, an attacker can use readily available tools to scan the Internet for potential victims, and can use similar tools to conduct brute force attacks.
[0002] The attackers targeted servers with relatively weak passwords and without multi-factor authentication, virtual private networks (VPNs), and other security protections. Through brute force attacks, the threat actor group has gained access to the target machines and has conducted a number of subsequent activities such as ransomware and coin mining operations. BRIEF DESCRIPTION OF THE DRAWINGS
[0003] Figure 1 is an overview diagram of an example embodiment of a network.
[0004] Figure 2A The empirical distribution of the number of days per machine on which one or more brute force attacks were observed is shown.
[0005] Figure 2B Shows sample counts of daily and maximum hourly failed network logons for the local machine under brute force attack.
[0006] Figure 3 Example daily and cumulative failed network logons are shown.
[0007] Figure 4A -B shows two histograms including example daily counts of Remote Desktop Protocol (RDP) public connections per machine in an example enterprise that was subject to a known brute force attack.
[0008] Figure 5 is a bubble chart of the example IP abuse score versus the count of machines with inbound RDP.
[0009] Figure 6 Shown are example daily counts of network logon failures for machines that were not subjected to brute force attacks.
[0010] Figure 7 An example ranking of detected RDP inbound brute force attacks is shown.
[0011] Figure 8 is an example data flow diagram of one or more disclosed embodiments.
[0012] Fig. 9 is an example data flow diagram of one or more disclosed embodiments.
[0013] Fig.10 is a flow chart of an example method for determining whether a network is under a brute force attack.
[0014] Fig.11 is a flow chart of an example process for estimating parameters of an unknown distribution.
[0015] Fig.12 The diagram illustrates a block diagram of an example machine 900 upon which any one or more of the techniques (eg, methodologies) discussed herein may be performed. DETAILED DESCRIPTION
[0016] Brute force attacks represent technical problems that can be difficult to detect when they generate a manageable number of false positives. In a brute force attack, an attacker can attempt to log into an account by effectively using one or more trial and error methods. In some cases of brute force attacks, failed logins are associated with these attacks. These failed logins may occur within a very short time frequency (usually a few minutes or even seconds). Brute force attacks may also involve attempts to access one or more accounts using valid usernames. An attacker can obtain a valid username via traditional credential theft techniques or by using a common username (such as "Administrator"). The same is true for password combinations. In detecting brute force attacks, since password data is not available in at least some cases, some embodiments of the present disclosure focus on the source Internet Protocol (IP) address and username.
[0017] In some operating environments, logon failures are recorded to a log file or other logging tool. These log entries may include an event code indicating a logon failure and associated information such as the username used in the logon attempt. For example, in the Microsoft Windows operating system, whenever a logon attempt fails for a local machine, Event Tracing for Windows (ETW) registers event ID 4625 with the associated username.
[0018] In addition, source IP addresses associated with failed login attempts are also recorded in at least some environments. This information can be useful in evaluating whether a machine is subject to a brute force attack. This information (combined with login failure indications for non-server machines (e.g., event ID 4624)) is useful for understanding which login sessions were successfully created. Based on this understanding, an indication of a compromised machine is generated.
[0019] Embodiments are disclosed that utilize these and / or other signals that have proven valuable in detecting and, in some embodiments, mitigating brute force attacks. In some embodiments, this capability can be deployed via Microsoft Threat Experts - a managed threat hunting service in Microsoft Guardian Advanced Threat Protection. Other embodiments are deployed in environments that include solutions provided by a mix of different vendors or in non-Microsoft environments.
[0020] In some embodiments, embodiments of the present disclosure generate one or more alarms. In various embodiments, the alarm is directed to one or more of a system administrator, a security vendor, an end user, or other contacts. In some embodiments, detection of abnormal activity indicating a brute force attack results in programmatic reconfiguration of network security devices such as firewalls, proxies, routers, switches, or other network devices to mitigate the brute force attack. For example, in some embodiments, based on the likelihood of a brute force attack, the access control policy of one or more network security devices is updated. In some embodiments, programmatic reconfiguration is performed without human intervention. For example, in some embodiments, a firewall is reconfigured to prevent a device identified as initiating at least a portion of the identified brute force attack from accessing the network. In some embodiments, stricter access policies are gradually deployed to network security devices until the brute force attack is successfully mitigated (e.g., the level of abnormal activity drops below a predetermined threshold). In some embodiments, these access policies are gradually deployed by programmatically configuring one or more firewalls deployed to control access to the network.
[0021] Some disclosed embodiments recognize that observing a sudden, relatively large number of events of a particular type (e.g., event ID 4625) associated with a network connection (e.g., RDP), although infrequent in some environments, does not necessarily mean that the machine is under attack. For example, a script that performs the following actions would appear suspicious when the count of failed logins over a time series is evaluated. However, this example is likely not malicious:
[0022] Using expired passwords
[0023] Retry login attempts every N minutes with a different username
[0024] Via a public IP address within a domain owned by the enterprise
[0025] In contrast, behaviors including the following events can indicate an attack:
[0026] Extreme counts of failed logins from many unknown usernames
[0027] Never successfully authenticated before
[0028] From multiple different connections (e.g., multiple different RDP connections)
[0029] From new inbound IP addresses (especially those with a history of abuse)
[0030] The above examples demonstrate that understanding the context of failed logins and inbound connections is useful in distinguishing between true positive (TP) and false positive (FP) brute force attacks. This understanding improves the accuracy of detecting brute force attacks.
[0031] In the development of embodiments of the present disclosure, historical data from previous months was obtained. This historical data was then analyzed to provide insights into the types of brute force attacks that occurred in the tested environment during these previous months. Among the approximately 45,000 machines with public IP connectivity (e.g., RDP) and at least one (1) failed network login, on average, hundreds of machines had a high probability of being subjected to one or more brute force attack attempts per day. Among the subset of machines where brute force attacks were detected, the attacks lasted an average of 2-3 days, with approximately 90% of the cases lasting 1 week or less, and less than 5% of the cases lasting 2 weeks or longer. The presently disclosed embodiments were further developed based at least in part on the insights gained through the analysis of the historical data.
[0032] To detect brute force attacks, some disclosed embodiments implement a technical solution to the technical problem of brute force attacks by modeling the distribution of operating parameter values as a finite mixture of distributions. The use of a mixture model in brute force attack detection can identify the multimodal nature of the operating parameter values experienced by a computer network. By comparing the operating parameter values to a distribution that is appropriate for the network environment within a given period of time, embodiments of the present disclosure adjust their detection of brute force attacks as needed to reduce the number of false positives while still providing accurate detection of true brute force attacks.
[0033] Each distribution in the hybrid model is defined by one or more parameters. The parameters, and therefore the distribution itself, are dynamically updated based on additional operating parameter values as these operating parameter values are observed. Embodiments of the present disclosure compare observed operating parameter values with specific distributions in the hybrid model based on a dynamically determined threshold that determines which distribution is used for the comparison. In some embodiments, multiple time series are each modeled as a separate mixture of distributions, each hybrid model being defined by its own distribution parameters and comparison threshold. By recognizing that one or more time series of operating parameter values may be multimodal, and therefore using a hybrid model of distributions for optimal comparison, embodiments of the present disclosure are able to more accurately detect abnormal events while reducing the probability of false positives.
[0034] Figure 1 is an overview diagram 100 of one or more networks in which embodiments of the present disclosure may be implemented. Figure 1A network 101 is shown. The network 101 includes a plurality of network components 102A-F. Figure 1 Also shown are devices 104A-C external to the network 101. The network components 102A-F within the network 101 and the external devices 104A-C outside the network 101 communicate with each other. An example of this communication is shown in FIG. Figure 1 This is illustrated as arrows 106A-G (not all arrows are labeled to maintain clarity of the drawing).
[0035] Access to network 101 is controlled by firewall 108 . Figure 1 Also shown is a network security system 110. In some embodiments, the network security system 110 collects information about communications between network components 102A-F and / or external devices 104A-C, represented by arrows 106A-G. In some embodiments, the network security system 110 also collects operating parameter values for one or more of the network components 102A-F. As discussed in more detail below, based on information indicating these communications, embodiments of the present disclosure provide an improved method of detecting brute force attacks on the network 101. In some embodiments, the network security system 110 sends one or more control signals 112 to the firewall 108. The control signals 112 configure one or more network access policies for one or more devices 102A-F.
[0036] The network security system 110 receives operational parameter values from one or more network components 102A-F. The operational parameter values are received by the network security system 110 via messages transmitted (directly or indirectly) from the network components to the network security system 110. These messages are shown as messages 122A-C. In at least some embodiments, the operational parameter values provided by one or more network components 102A-C are provided as a time series of values. In other words, the operational parameter values are measured periodically or multiple times at regular or irregular intervals and then communicated to the network security system 110.
[0037] Network security system 110 processes the operating parameter values according to the methods discussed herein to determine the probability that network 101 is experiencing a brute force attack. For example, in some cases, analysis performed by network security system 110 determines that one or more external devices 104A-C are launching a brute force attack on network 101.
[0038] Figure 2A Figure 1 shows an example empirical distribution of the number of days a network device experiences a brute force attack. A large number of failed logins is often associated with a brute force attack. Figure 2A In the example, 90% of brute force attacks exhibit more than 10 attempts, with the median being greater than 60. Additionally, unusual daily counts have a high positive correlation with large counts in shorter time windows (see Figure 2B). The daily extreme number of failed logins typically occurs within two hours, with approximately 40% of failures occurring within 30 minutes.
[0039] Figure 2B Example counts of daily and maximum hourly network failed logins for a local machine under brute force attack are shown. While many brute force attacks can be detected based on detection logic that thresholds the failed login counts over a daily or finer-grained time window, this can generate too many false positives. Relying on such a strategy will also result in false negatives, leading to network leakage. We identified several instances of brute force attacks that generated fewer than five to ten failed attempts per day but often lasted for multiple days. This attack pattern avoids extreme counts at any point in time. For this brute force attack, as described below Figure 3 As discussed, thresholding the cumulative count over time of unsuccessful logins across time demonstrates effectiveness.
[0040] Figure 3 Daily and cumulative failed network logons are shown. Viewing the count of failed network logons provides a useful but incomplete picture of RDP brute force attacks. This can be further enhanced with additional information about the failed logons, such as the reason for the failure, the time of day and day of the week, and the username itself. A particularly strong signal is the source IP of the inbound RDP connection. Knowing whether the external IP has a high abuse reputation (which can be found on sites such as https: / / www.abuseipdb.com / ) can directly confirm whether the IP is part of an active brute force attack.
[0041] Unfortunately, not all IP addresses have a history of abuse, and retrieving information about many external IP addresses on demand can be expensive. Maintaining a list of suspicious IPs is an option, but relying on this can lead to false positives as new IPs inevitably appear all the time (especially with cloud computing and the ease of launching virtual machines). A common signal that can enhance failed logins and user information is counting distinct RDP connections from external IP addresses. Likewise, extreme values occurring at a given time or accumulated over time can be indicators of an attack.
[0042] Figure 4A and Figure 4B An example histogram of daily counts of RDP public connections per machine (i.e., counts put into discrete bins) is shown for an example enterprise that was subject to a known brute force attack. Clearly, a healthy machine has a lower probability of having a larger count than an attacked machine.
[0043] Given that some enterprises have machines that are attacked by brute force on a daily basis, priority can be focused on compromised machines, which are defined by the first successful login after an unsuccessful attempt from a suspicious source IP address or unusual username. In an embodiment for the Microsoft Windows operating system, the Windows log, event ID 4624, can be leveraged to measure successful login events for the local machine in conjunction with unsuccessful logins (e.g., event ID 4625).
[0044] Of the hundreds of machines that were detected as being compromised by RDP brute force attacks during our analysis, we found that approximately 0.08% of them were compromised. Furthermore, across all enterprises analyzed over a period of several months, an average of approximately one machine was detected being compromised every 3-4 days due to RDP brute force attacks with a high probability.
[0045] Figure 5 A bubble chart showing example average abuse scores for external IPs associated with RDP brute force attacks that successfully compromised machines. The size of the bubble is determined by the count of different machines in the enterprise being analyzed that have network connections from each IP. While the IP sources are diverse, the Netherlands, Russia, and the United Kingdom have a large concentration of inbound RDP connections from high-abuse IPs.
[0046] From our analysis, we concluded that successful brute force attempts are not uncommon; therefore, it is critical to at least monitor for suspicious connections and unusual unsuccessful logins that lead to verified login events. In the following sections, we describe methods for doing so.
[0047] Reliance on thresholding on operational parameter values (such as the number of unsuccessful attempts per machine for detecting brute force attacks) can be noisy and can result in many false positives. For example, in a scenario where a script has been preconfigured with expired passwords and continually attempts to log in with invalid passwords, this approach may generate false positives.
[0048] To avoid the problems associated with basic thresholding, at least some disclosed embodiments utilize multiple context-dependent signals, such as:
[0049] Timing, type and count of unsuccessful login attempts
[0050] Username History
[0051] Type and frequency of network connections
[0052] · A first-time user name for a new source machine with a successful login
[0053] These examples can be expanded to include indicators associated with brute force attacks, such as counts of detected port scans.
[0054] When determining whether an RDP inbound brute force attack is ongoing or has previously occurred, at least some disclosed embodiments analyze one or more of the following per-machine (or device) signals:
[0055] Unsuccessful Logins and RDP Connections: Time of day and day of week. In various embodiments, this data is represented via a time series indicating counts of unsuccessful logins and / or a separate time series indicating counts of successful logins. These time series are associated with the time identifying counts of unsuccessful or failed logins per hour and / or day as needed. In various embodiments, these time series are generated on a per-device and / or per-user / source IP basis. In some embodiments, these time series can be aggregated across time periods. For example, a time series representing weekly impacts can be generated (e.g., in some cases, activity from Monday to Wednesday may be greater than end of work week or weekend activity). Some embodiments generate time series that accumulate counts across several weeks.
[0056] Time delays between one or more failed or unsuccessful logins and successful logins. For example, in some embodiments, a time series is generated indicating one or more of user account information such as user account name, IP address, last failed login time, and last successful login time.
[0057] Event ID 4625 Logon Type (filtered to network and remote interactions). This time series provides logon type information for each logon attempt and / or successful logon. In some embodiments, the logon attempt is a Remote Desktop Protocol (RDP) logon attempt. In some embodiments, the time series is an ordered time series. In some embodiments, the logon type information (such as logon type) is represented using a standard logon type code, such as code=3 for a network logon, code=10 for a remote logon, and code=1 for a local console initiated logon.
[0058] Event ID 4625 Failure Reason (filtered as %%2308, %%2312, %%2313). Some embodiments generate a categorized time series indicating the code for the reason for the login failure. For example, %%2313 is the code for "unknown username or bad password".
[0059] A cumulative count of each different or unique username that failed to log in without a successful login (or without a subsequent successful login). Some embodiments include one or more time series related to a single device or across multiple devices. For example, in some embodiments, a first time series conveys a count C of different usernames that attempted to log in at a particular device and failed. A second time series conveys a second number "k" of C different users that also failed to log in to other devices.
[0060] · Count (and cumulative count) of failed or unsuccessful logins. As discussed above.
[0061] · Counts (and cumulative counts) of RDP inbound external IP addresses. In some embodiments, the time series provides different counts, external IP addresses per device over time, per device per IP address, and / or per IP address (which may include IP addresses across enterprise networks). Some embodiments include additional time series of counts of specific IP addresses across devices over time. In various embodiments, these time series are for each machine / device, and / or across devices. For example, in some embodiments, user and IP address statistics are checked at each machine and across machines. In some embodiments, counts of failed / successful / network connections are checked on a per-machine basis. Some embodiments include a time series of cumulative counts, which, in addition to providing counts for longer periods, also provide counts of similar data as described above. For example, in some embodiments, a first time series provides counts for a period, where the period is specified via a configurable parameter (e.g., every hour). These embodiments can also provide a second time series of cumulative counts including parameters for a longer period (e.g., 12 hours, 24 hours, one day, two days, three days, one week, or any other longer period).
[0062] Counts of other machines that have RDP inbound connections from a common device or at least devices with equivalent IP addresses. In some embodiments, this time series is related to a count of IP addresses that have attempted to log in on multiple devices. Described another way, the time series indicates a count of at least two devices that have RDP inbound connections from a common device. In at least some embodiments, these time series counts are generated for each device included within the monitoring network. Thus, in these embodiments, for a particular device, the count represents a count of devices with inbound connections that have a source IP address that is equivalent to the source address of the inbound connections on that particular device.
[0063] In some embodiments, each time series or signal is scored for its indication of a brute force attack. In some embodiments, the hourly time window is the lowest time granularity. In other embodiments, other time periods (one minute, five minutes, etc.) are used as the shortest time window for the time series. One or more additional time series can be generated by accumulating the time series of the lowest time granularity. For example, in some embodiments, some time series aggregate multiple hourly time series within a day. In some embodiments, some time series aggregate granular time series across multiple days to take into account specific daily and weekly periodic effects. In addition, in some embodiments, some time series are aggregated across multiple devices and across multiple networks.
[0064] Let t be a time window and T be the interval of multiple time windows with a time frequency higher than t. The following discussion will be t is defined as the lowest level univariate time series for each signal type and each single or multiple devices, and y T is y of time frequency T t Aggregation.
[0065] The first signal is a categorized time series of login types per device at each t (e.g., event ID 4625). For each device in the network, y t =[("2", count t ), (″3″, count t )...,(″11″,count t )], where "2", "3", and "11" are examples of logon types. For example, "2" is a logon at the keyboard, "3" is a connection to a shared folder on the device from elsewhere on the network, and so on. Of particular interest for monitoring inbound brute force are logon types "3" and "10". count t Represents the total number of failed logins at time t for a specified login type. In some embodiments, the count is determined across groups of usernames and IP addresses.
[0066] The second signal is a categorized time series of failure reasons for each login type, each t, and each device (e.g., event ID 4625). For each device in the network, y t =[(″2:%%2309″,count t ), ("3:%%2313", count t )..., ("11:%%2304", count t )], where the time series y t Each entry includes a login type (e.g., "2:", "3:", "11:"), a failure reason code (e.g., "%%2309"), and a count. For example, the failure reason %%2308 describes that the user attempting to log in is not authorized to access the device. t The term represents the total number of failed logins at time t for a specified login type and failure reason. In some embodiments, the total number of failed logins is counted across usernames. Some embodiments generate a time series count for each IP address t In some embodiments, an aggregate count is generated for a group of IP addresses having an associated reputation or anomaly score that meets a predetermined criterion (eg, a reputation or score less than a predetermined threshold). t Time series of values.
[0067] In some embodiments, the third signal is a count of failed logins aggregated over a time series across multiple login types and / or multiple login failure reasons. Thus, these embodiments contemplate a third signal comprised of one or more combinations of possible login types and one or more combinations of possible login failure reasons. In some embodiments, the signal is generated on a per-device basis. For each device in the network and the set of login type and failure reason time series, y t = count of the total number of failed logins at time t across usernames and IP addresses. For example, y t This could be a count summed across possible logon types 3 and 10 and logon failure reasons in (%%2308, %%2312, %%2311, %%2313).
[0068] In some embodiments, the fourth signal is a cumulative count of failed logins per device per T. The cumulative count is filtered by one or more login types and login failure reasons. For each device in the network and the filtered login type(s) and failure reason(s), the time series y T =∑ t∈T y t , where y T is the count of failed login signals.
[0069] In some embodiments, the fifth signal is a cumulative count of failed logins per device per T across a subset of one or more login types and failure reasons. For each device in the network and a set of time series of login types and failure reasons, y T =∑ t∈T y t , where y T is a count of failed login signals. Thus, the difference between the fourth and fifth signals is that the fifth signal includes a subset of login types relative to the fourth signal. For example, in some embodiments, the fifth signal indicates a login failure with a reason associated with a new (or unknown) username and a login type indicating a network connection (e.g., the login attempt was initiated from outside the monitored network or at least not via the console of the machine).
[0070] In some embodiments, the sixth signal is a total count of inbound common source addresses of the Remote Desktop Protocol (RDP) for each device at time t. In other words, in some embodiments, the sixth signal indicates how many different source addresses attempted RDP connections on the device during time period t. Thus, the sixth signal provides a common RDP source address for each device in the network. Thus, in some embodiments, the sixth signal may be defined as y T =Total count of different IP source RDP addresses per t.
[0071] In some embodiments, the seventh signal is a cumulative count of RDP inbound public source addresses for each device. For each device and RDP network connection in the network, y T =∑ t∈T y t , where y t is the count of RDP inbound connections. The seventh signal accumulates counts for a longer period of time than the sixth signal. For example, although in some embodiments the sixth signal accumulates counts for a period of one hour, the seventh signal accumulates counts for a period of one day, two days, three days, or even longer.
[0072] In some embodiments, the eighth signal is the cumulative failed and successful login rate for each username per time period t. In some embodiments, the eighth signal is generated for each device and for each username used in the attempted login. Therefore, in some embodiments, let y T =(y T,fail ,y T,success ) is the binary cumulative number of failed and successful logins accumulated across all devices in the network, across each time window up to T. Let U T is the count of distinct usernames on the device, where y T,success =0, i.e., the user has never been successfully authenticated to any device in the network. Therefore, in some embodiments, the eighth signal represents a time series of cumulative counts of usernames that have attempted to log into the device but have never been successful. In at least some embodiments, the eighth signal is generated,
[0073] In some embodiments, the ninth signal is the last successful login of the username in each T. For each device and each username, let y T =(T1, ..., T k ) are the k most recent different times of at least one successful login. If it does not exist, this time series is set to empty.
[0074] In some embodiments, the tenth signal characterizes the IP address abuse per T. For each IP address (if applicable), let y T-d is the reputation score of the IP address associated with suspicious activity (brute force, scanning, etc.) during period T and the last d days. Otherwise, y T-d is a multivariate time series consisting of:
[0075] ·The region associated with the source IP address. Some embodiments assign reputation scores to geographic regions. In some embodiments, the reputation of the source IP address is determined based at least in part on the reputation associated with the origin region. In these embodiments, the second component of the reputation is based on the source IP address itself. In some embodiments, if the history associated with the source IP address is less robust, the region of the source IP address has a greater weight. In other words, if the reputation information associated with the source IP address meets the first criterion, the first weight is assigned to the reputation of the source IP address region, and the second weight is assigned to the reputation of the source IP address itself. Then, the overall reputation is determined based on the first weight and the second weight. If the source IP address reputation meets the second criterion, the third weight is assigned to the reputation of the source IP address region, and the fourth weight is assigned to the reputation of the source IP address itself. In some embodiments, the third weight is lower than the first weight. In some embodiments, the fourth weight is higher than the second weight.
[0076] Counts of devices with multiple inbound RDP connections and failed logons.
[0077] A count of enterprises with at least one device having multiple inbound RDP connections and failed logons.
[0078] A count of distinct IP addresses that also occurred during RDP connections with failed logons within the time window on a given device
[0079] For many cybersecurity problems, including detecting brute force attacks, previously labeled data is often not available. Therefore, training a supervised learning model is not feasible. This is where unsupervised learning helps, enabling one to discover and quantify unknown behaviors when examples are too sparse. Given that several of the signals we consider for modeling RDP brute force attacks are inherently dependent on values observed over time (e.g., daily counts of failed logins and counts of inbound connections), time series models are particularly beneficial. Specifically, time series anomaly detection naturally provides a logical framework to quantify uncertainty in modeling temporal variation in the data and produce probabilities that can then be sorted and compared to predetermined thresholds to control the desired false positive rate.
[0080] Time series anomaly detection captures the temporal dynamics of signals and accurately quantifies the probability of an observed value at any point in time under normal operating conditions. More formally, if we introduce the notation Y(t) to denote the signal taking a value at time t, then we build a model to compute a reliable estimate of the probability that Y(t) exceeds the observed value given all known and relevant information, denoted by P[y(t)], sometimes referred to as the anomaly score. Given a false positive tolerance rate r (e.g., 0.1% or one in ten thousand for each time step), for each time t, a value y*(t) that satisfies P[y*(t)] < r will be detected as an anomaly. Assuming the correct signals reflecting relevant behaviors of the attack type are selected, the idea is straightforward because the lowest anomaly score occurring for each time step is likely associated with the highest likelihood of a true threat.
[0081] For example, regarding the Figure 2B discussed above, the time series of the daily count of failed logins on the day of the brute force attack on August 4, 2019, has extreme values that are associated with an empirical probability of approximately 0.03% among all machines and days with at least one failed network login for the enterprise.
[0082] As previously discussed, applying anomaly detection to a single or a few signals to detect true attacks may generate too many false positives. To mitigate this, we combine the anomaly scores across eight signals we selected to model the RDP brute force attack pattern. Details of our scheme are included in the appendix, but in general, our approach includes:
[0083] 1) Sequentially update a statistical discrete time series model for each signal, capturing the time of day, day of the week, and point and cumulative effects
[0084] 2) Use methods that produce accurate probability estimates to combine the anomaly scores, and
[0085] 3) Rank the top N anomalies per day to control the expected number of false positives. Our method for time series anomaly detection is computationally efficient and automatically learns how to update probabilities and adapt to data changes.
[0086] Figure 6 shows the daily count of network failed logins for machines without brute force attacks. As Figure 6 can be seen, the parametric discrete location / scale distribution does not generate well-calibrated p-values for rare time series and thus may result in too many false positives when used to detect anomalies when looking at many machines at a high time frequency. To overcome this challenge for sparse time series of counts of failed login attempts and RDP inbound public connections, some embodiments utilize a mixture model. In some of these embodiments, a zero-inflated two-component negative binomial distribution is used.
[0087] The formula is based on a thresholding value that is used to select a distribution to compare with the time series. An example of selecting a distribution is provided via Equation 1 below. The threshold-based comparison of a single time series to multiple distributions captures multimodal and heavy-tailed behavior of operating parameter values in a computationally efficient manner. This represents an improvement over other methods such as those utilizing expectation maximization.
[0088] In some embodiments, the stratification prior is given by empirical estimates of sample moments on a machine using approximately one month of data.
[0089] Equation 1: Zero-adjusted negative binomial threshold model
[0090] Let y t is a univariate time series corresponding to one of the violence signals.
[0091]
[0092] Stratified Priors:
[0093]
[0094]
[0095]
[0096]
[0097] Negative Binomial Distribution (NB):
[0098] Given estimates μ = mean(y), and σ = standard deviation(y).
[0099] The calculation scale s = (σ^2-μ) / (μ^2).
[0100] PDF: P(Y= y|μ , s)=『(y+1 / s) / 『(1 / s)『(y+1)((s*μ) / (1+s*μ))^y(1 / (1+s*μ))^{1 / s}
[0101] in is the gamma function 4
[0102] This formula does not produce a conjugate prior, so the probability cannot be calculated directly from the posterior predictive density. Instead, the anomaly score is generated based on drawing samples from the distribution and then calculating the empirical right-tail p-value.
[0103] Updating the parameters is done based on applying exponential smoothing, which is a sequential and computationally efficient method that weights more recent data. To avoid biasing estimates based on anomalous data (such as machines being subjected to brute force or other attacks), trimming is applied to samples from the distribution with a specified false positive rate (e.g., 0.1%). Algorithm 1 below demonstrates this.
[0104] The smoothing parameters are learned based on maximum likelihood estimates and then fixed during each new sequential update. To induce further uncertainty, bootstrapping across machines is done to produce a histogram of smoothed weights, and samples are drawn based on their frequency. We find that for over 90% of the machines, the weights in the self-0 set vary between 0.06% and 8%, resulting in slow parameter changes. Extensions using adaptive forgetting factors will be considered in future research to automatically learn how to correct for smoothing in real time.
[0105] To update the model parameters, some embodiments utilize Algorithm 1, as discussed below. In Algorithm 1, for each model parameter θ t , smoothing parameter distribution α θ and the new value y t+∈ , perform the following steps:
[0106] 1) Apply an appropriate link function g θ To generate a continuous unbounded variable η t :=g θ [θ t ];
[0107] 2) Sampling smoothing weights
[0108] 3) Use exponential smoothing to update additive parameters:
[0109]
[0110] 4) Update parameter θ t+∈ =g θ -1 [η t+∈ ].
[0111] Figure 7 An example ranking of detected RDP inbound brute force attacks is shown. For each machine detected as potentially vulnerable to a brute force attack, each instance is assigned a TP, FP, or unknown, and each TP is assigned a priority based on the severity of the attack. For high priority TPs, in some embodiments, a targeted attack notification with detailed information and recommendations about the active brute force attack is sent to the associated enterprise; otherwise the machine is closely monitored until more information is available.
[0112] Some disclosed anomaly detection embodiments provide the additional ability to send targeted attack notifications to an organization when an attack is detected. In many cases, these notifications are provided before a brute force attack is successful or the actor is able to conduct further malicious behavior. Experimental results show that the daily average precision (i.e., correct positive rate) is approximately 93.7% at a conservative false positive rate of 1%.
[0113] By selecting signals that were found to be highly correlated with RDP brute force attacks, time series anomaly detection was highly accurate in identifying real threats. Monitoring for suspicious activity in failed logins and network connections should be taken seriously, and real-time anomaly detection that can dynamically self-update as the network changes can indeed provide a sustainable solution.
[0114] Some embodiments use a hierarchical zero-adjusted negative binomial dynamic threshold model to capture the characteristics of highly discrete count time series. Specifically, Figure 2B As shown in , many examples of brute force attacks do not demonstrate failed logins with valid credentials on the local machine, so there are excess zeros that cannot be explained by standard probability distributions (such as negative binomial). In addition, the variance of the non-zero counts can be much larger than the mean, where, for example, a valid script connecting via RDP may generate 20 or more counts over several minutes due to password expiration. Furthermore, given a combination of multiple users or scripts connecting to a shared machine simultaneously, this can produce more extreme counts at higher quantiles, leading to examples such as Figure 6 The heavier tail is visible in the figure.
[0115] Figure 8 is a diagram that illustrates the flow of data that occurs in one or more disclosed embodiments. Figure 8 The data flow 800 shows the operation of two separate hybrid models operating on two separate time series 802A and 802B. Each model includes multiple distributions (e.g., 804A-C in hybrid model 1 and 804D-F in hybrid model 2). Each hybrid model also has thresholds that define which of the multiple distributions is applied to the time series values (shown as value domains 805A-C and domains 805D-F). Each hybrid model also includes a threshold hybrid component that compares the current time series value to the selected distribution. Each hybrid model generates an anomaly indication, such as Figure 8 These anomaly indications are then combined by combiner 814 to generate a single anomaly indication or anomaly score 816.
[0116] In particular, data stream 800 includes a plurality of time series 802A and 802B. Each of time series 802A and 802B indicates an operational parameter value. Each of time series 802A and 802B is provided to selection components 803A and 803B, respectively. Selection components 803A-B select between a plurality of distributions based on the operational parameter values of the respective time series and a domain associated with each of the plurality of distributions. For example, Figure 8 Value ranges 805A-C are shown that are associated with distributions 804A-C, respectively. Figure 8 Also shown are domains 805d-f associated with distributions 804d-f, respectively. Selection component 803A selects one of distributions 804A-C of time series 802A based on value domains 805A-C. Selection component 803B selects one of distributions 804D-F of time series 802B based on domains 805D-F. Value domains 805A-C and 805D-F are dynamically updated by some embodiments of the present disclosure. For example, some embodiments update value domains 805A-C and 805D-F based on the values of time series 802A and 802B, respectively. This dynamic update process will be discussed further below.
[0117] The selected distributions (809A for time series 802A, and 809B for time series 802B) are provided to the corresponding threshold mixture models 808A-B. In some embodiments, each distribution P[Y|⊙] is modeled as a finite mixture of distributions. The mixed distributions are selected from the distribution list P1[Y|⊙1], ... Pm[Y|⊙ m ], where each parameter ⊙ p is a random process.
[0118] Each threshold mixture model 808A-B compares the corresponding time series with its selected distribution (809A for time series 802A and 809B for time series 802B) and outputs an anomaly indication, shown as 812A-B, respectively. Anomaly indications 812A-B are combined by combiner 814 into a single anomaly indication or anomaly score 816. In some embodiments, combiner 814 combines anomaly indications 812A-B using the Fisher method as discussed further below. In some embodiments, single anomaly indication or anomaly score 816 indicates whether the network is under a brute force attack.
[0119] Although Figure 8Two time series, two distribution groups (e.g., 804A-C and 804D-F), two selectors (803A-B), and two threshold mixture models (e.g., 808a-b) are shown, but some disclosed embodiments rely on more than two of each of these components. For example, as discussed above, any number of time series (e.g., input signals), such as three (3), four (4), five (5), ten (10), fifteen (15), twenty (20), or any number can be processed, compared, and combined (e.g., via combiner 814) in contemplated embodiments.
[0120] Figure 8 Two distribution groups are shown as 804A-C and 804D-F. Figure 8 802A, and a second selection of one of the distributions 804D-F for comparison with the time series 802B. In some embodiments, the two distribution groups 804A-C and 804D-F are equivalent. In some embodiments, the two distribution groups 804A-C and 804D-F overlap. In some embodiments, the two distribution groups 804A-C and 804D-F overlap but are not equivalent. This may also be the case for more than two distribution groups, such as three (3), four (4), five (5), ten (10), twenty (20), or any number of distribution groups.
[0121] Figure 8 Updater 851 is also shown. Updater 851 monitors time series 802A or time series 802B (which is not in Figure 8 As shown, updater 851 is shown as dynamically updating the hybrid model for time series 802A based on the monitored operational parameter values. In particular, updater 851 is shown as generating new parameter values ⊙ for each of the distributions 804A-C included in a particular hybrid model. These parameter values are Figure 8 The updater 851 also updates the thresholds that define the boundaries between the distributions of the mixture model. For example, the thresholds define the boundaries between the distributions 804A-C. These boundaries are Figure 8 It is represented by different value ranges 805A-C. Figure 8 The updated value of τ is shown via data stream 824 defining the value range 805A-C. Updater 851 also updates the updated value of τ by data stream 822. Figure 8 The symbol “C” in denoted a two-state homogeneous hidden Markov model. Figure 8Updater 851 is shown as being applied only to the top hybrid model that is processing time series 802A, but in at least some embodiments, the updater is similarly applied to two or more hybrid models used to detect brute force attacks. For example, updater 851 can also be applied to update a second hybrid model that is processing time series 802B, however, this has been changed. Figure 8 The specific method used by the updater 851 will be discussed in more detail below.
[0122] Fig. 9 An example data flow implemented in one or more disclosed embodiments is shown. Fig. 9 Data flow 850 illustrates how updater 851 processes data in one or more disclosed embodiments. Fig. 9 A time series of operating parameter values 852 flowing into a historical database 854 is shown. The historical time series data 857 is then provided to an updater 851. Based on the historical time series data, the updater generates data values 858 (τ), 860 (C), and 862 (⊙). Data value 862 is a parameter that at least partially defines the distribution of the historical time series data 857. In some embodiments, multiple parameter values are generated. Data value 860 is a two-state homogeneous hidden Markov model defined by the probability that the time series data (e.g., 857) is greater than the value τ of data value 858. Return to reference Figure 8 , data values 858, 860, and 862 generated by updater 851 are shown flowing to control value ranges 805A-C, distributions 804A-C, and threshold mixture models 808A-C.
[0123] Fig. 9 The τ value in data value 858, the two-state Markov model value in data value 860, and the parameter value in data value 862 stored in the hybrid model parameter data store 864 are shown. In some embodiments, the hybrid model parameter data store 864 is referenced by the scoring process 866. Figure 8 As discussed, scoring process 866 implements at least a portion of data flow 800 .
[0124] In at least some embodiments, the updater 851 operates on multiple time series and multiple corresponding mixture models obtained by the disclosed embodiments. One implementation of the updater 851 is described below with respect to Fig.11 Have a discussion.
[0125] Fig.10 is a flow chart of a process for determining whether a network is under a brute force attack. Fig.10One or more functions discussed are performed by hardware processing circuitry (e.g., 1202). In some embodiments, instructions (e.g., 1224) stored in memory (e.g., 1204, 1206) configure the hardware processing circuitry to perform one or more functions discussed below. In some embodiments, process 1000 discussed below is performed by a network management system or network security system (e.g., 110).
[0126] After start operation 1005, process 1000 moves to operation 1010, where a first time series of operating parameter values is received. The first time series of operating parameter values is related to the operation of the device attached to the network. For example, as described above with respect to Figure 1 As discussed, the operational parameter values of one or more network components 102A-F are provided to the network security system 110. In some embodiments, operation 1010 includes receiving one or more messages (e.g., any one or more messages 122A-C) indicating the operational parameter values. In various embodiments, multiple different time series of different operational parameter values are received. In some embodiments, multiple different time series are received from a single device and / or multiple devices. For example, as discussed above, one or more network component devices provide or transmit time series indicating operational parameter values to a network management system or network security system (e.g., 110). The received time series of operational parameter values indicates one or more failed logins and RDP connections for each hour of the day and each day of the week, the time delay between one or more failed logins and successful logins, the login type (e.g., event ID 4625), the failure reason (e.g., event ID 4625), the cumulative count of each different username of unsuccessful failed logins, the count (and cumulative count) of failed logins, and the count (and cumulative count) of IP addresses outside of RDP (e.g., for inbound RDP connections). The time series of these operational parameter values are discussed in more detail above.
[0127] In operation 1020, the operating parameter values in the first time series are compared with the first parameter value range and the second parameter value range. For example, as discussed above, some disclosed embodiments employ a multimodal distribution (mixture model) approach to analyze the operating parameter values.
[0128] For example, the multiple time series of operating parameter values received in operation 1010 may be transmitted via y1 . . . y k Each time series is derived from the same given data generating process P[y i |⊙ i ], where each P[.|⊙ i ] can be different probability distribution families. Each distribution P[Y|⊙] is modeled as a distribution derived from distributions P1[Y|⊙1], ...P m[Y|⊙ m ] is a finite mixture of distributions specified by a list of p is a random process.
[0129] Some disclosed embodiments utilize a switching mechanism based on a learned threshold of y exceeding a specified quantile. In at least some embodiments, the quantile is defined via the following equation 1:
[0130] Let q p ∈(0,1) is P p [Y|Θ p ] is the qth quantile Q selected p , and define τ p : =E[Y|Y>Q p-1 ], so that τ p >τ p+1 .
[0131] Let C p It is a two-state homogeneous hidden Markov model, with Y>=τ p is defined by the probability of .
[0132] In some embodiments, the dynamic threshold mixture density is defined as follows:
[0133] P thesh [Y|Φ=(Θ,τ,C)]:=P1[Y|Θ1]*P[C1]if Y<=τ1;
[0134] P2[Y|Θ2]*P[C2]if τ1<Y<=τ2; (1)
[0135] ..., P m [Y|Θ m ]*P[C m, ]otherwise;
[0136] Thus, in operation 1030, a distribution is selected from a plurality of possible distributions to be applied to the time series based on the parameter range of the time series. Fig.11 ), in some embodiments dynamically adjusting the parameter range of each possible distribution. Therefore, in operation 1030, based on the comparison, a first distribution is selected from the plurality of possible distributions. For example, in at least some embodiments, if the time series values are generally within the first parameter range, the first distribution is selected, and if the time series values are generally within the second parameter range, the second distribution is selected. If the first distribution is selected, the selected first distribution is used to determine whether the time series represents abnormal activity or the degree of abnormal activity present in the time series.
[0137] In some embodiments, a time series value is generally within a first parameter value domain if the time series value falls within the first parameter value domain after outliers or exceptional values are eliminated or not considered. For example, some embodiments may determine the percentage of time series values that fall within each possible parameter value domain, whose domain includes the highest percentage of values for selecting an appropriate distribution.
[0138] In operation 1040, the first time series is compared to the selected distribution. For example, in some embodiments, a percentage of operating parameter values that conform to the selected distribution is determined. In embodiments where multiple time series are processed, the values of each time series are compared to a corresponding distribution selected based on the corresponding time series (e.g., in Figure 8 , the selected distribution 809A is compared with the time series 802A, and the selected distribution 809B is compared with the time series 802B).
[0139] As discussed above, process 1000 compares the values of one or more time series (e.g., time series 802A and time series 802B) to a corresponding set of parameter value ranges (e.g., value ranges 805A-C and value ranges 805D-F, respectively), or a first parameter value range and a second parameter value range for that particular time series. A distribution specific to each time series is then selected based on which range the values of the corresponding time series fall into (e.g., Figure 8 As illustrated in FIG. 8 , consider selecting multiple distributions 804A-C based on corresponding value ranges 805A-C and values of time series 802A).
[0140] In operation 1050, an anomaly indication is determined based on a comparison of the first time series with the selected distribution (e.g., performed when the threshold mixture model 808A compares the time series 802A with the selected distribution 809A). Some embodiments determine the percentage of the operating parameter values of the first time series that fall outside the selected distribution. The percentage is an anomaly indication. In other embodiments, the probability of occurrence of each value in the time series is determined based on the selected distribution. These probabilities are then aggregated to determine the anomaly indication. In some embodiments, the aggregation is performed by multiplying the probability of occurrence of each value. The resulting aggregated probabilities are then normalized in some embodiments.
[0141] Note that some embodiments of process 1000 operate on multiple different time series of operating parameter values and multiple mixture models. Figure 8As described, some embodiments combine multiple anomaly indications (e.g., 812A and 812B) generated by corresponding multiple hybrid models via a combination process (e.g., combiner 814) to generate a combined anomaly indication, or in other words, a single anomaly indication or anomaly score (e.g., 816). Some embodiments utilize Fisher's method to combine multiple anomaly indications into a single anomaly indication. For example, some embodiments operate as follows: Let α T,1 , ..., α T,N is the set of available scores. These embodiments then calculate where χ T is an anomaly indicator combined for the purpose of discussing process 1000. Fisher showed that if each α T,i are independent and uniformly distributed between 0 and 1, then χ T The chi-square distribution will have 2N degrees of freedom. Applying the Fisher test to the anomaly scores yields a scalable scheme that produces interpretable probabilities, which can therefore be controlled to achieve a desired false alarm rate.
[0142] At operation 1060, based on the anomaly indication, a likelihood of a brute force attack on the network is determined. For example, in some embodiments, if the anomaly indication is above a predetermined threshold, the process 1000 determines that the network is under a brute force attack. Some embodiments maintain a moving average of the anomaly indication, and if the moving average of the anomaly indication exceeds a predetermined threshold, it is determined that the network is under a brute force attack.
[0143] When it is determined that a brute force attack may be occurring, additional forensic processing is performed in some embodiments to identify one or more culprit computers that are performing the brute force attack. These identified computers are located outside the network (e.g., external devices 104A-C), or in some embodiments may be inside the network (e.g., in some cases, malware installed on computers within the network may provide a "proxy" for a malicious actor to perform brute force attacks on the network).
[0144] In operation 1070, brute force attacks are mitigated based on likelihood. Some embodiments mitigate brute force attacks by determining mitigation actions based on likelihood. In some embodiments, mitigation actions may include controlling network access. For example, in some embodiments, if the likelihood is above a predetermined threshold, one or more access control policies of a firewall (e.g., firewall 108) or multiple firewalls that control access to the network (e.g., 101) may be adjusted. For example, a stricter firewall policy (e.g., initiated by network security system 110) may be applied during a brute force attack. In some embodiments, packets received from devices identified as participating in a brute force attack (e.g., any one or more of external devices 104A-C) are dropped at a network access point (e.g., one or more firewalls, such as firewall 108). In some embodiments, this is achieved by modifying the access policy of the firewall to restrict access to the network by the identified device. In some embodiments, one or more alarms are generated based on likelihood. In various embodiments, an alarm may be generated via text, email, or other messaging technology.
[0145] In operation 1080, based on the operating parameter values of the first time series (e.g., 802A), the first parameter range and the second parameter range (e.g., range 805A and range 805B) are updated. Updating the first parameter range and the second parameter range at least adjusts the boundary between the first parameter range and the second parameter range. To the extent that the hybrid model includes more than two distributions, multiple boundaries are adjusted in operation 1080.
[0146] For example, given a sample y from P[Y|Φ] (e.g., a first time series), at least in some embodiments, the estimation of the parameter Φ is performed in two stages. In the first stage, the parameters are initialized. In some embodiments, the parameters are initialized using Gibbs sampling. It calculates the conditional likelihood of each parameter under the choice of a specified stratified prior. In some embodiments, the first stage is performed before performing operation 1010. In the second stage, the parameters are initialized using a stratified prior for the model with a link function g. Φ The smoothing weight α of each parameter Φ Φ The generalized exponential smoothing of the chosen grid is updated for each parameter. Therefore, for each parameter Φ, the link function g Φ is applied to produce the additive term η t :=g Φ[ Φ t] Then, the smoothing weights are sampled The additive parameters are updated according to the following formula:
[0147]
[0148] Among them, MΦ is the central moment corresponding to the parameter Φ.
[0149] In some embodiments, the updated parameter is set to Φ t+∈ =g Φ -1 [η t+∈ ]. After the distribution parameters are updated, the updated posterior distribution Φ t+1 , and for each new sample y t+1 , calculate the right tail p value. In some embodiments, the right tail p value is calculated according to the following formula:
[0150]
[0151] in:
[0152] Π t+1 is the right-tail p-value of the updated distribution,
[0153] μ is a probability measure; for discrete data, it is a count measure, otherwise it is a Lebesgue measure,
[0154] φ t+1 is the updated posterior distribution,
[0155] y t+1 is the new sample value,
[0156] P thresh is the dynamic threshold mixture density. In some embodiments, P thres is defined as
[0157] P thesh [Y|Φ=(Θ,τ,C)]:=P1[Y|Θ1]*P[C1]if Y<=τ1;
[0158] P2[Y|Θ2]*P[C2]if τ1<Y<=τ2;
[0159] ..., P m [Y|Θ m ]*P[C m, ]otherwise;
[0160] in:
[0161] Cp is a two-state homogeneous hidden Markov model, where Y>=τ p The probability definition of
[0162] τ p ::=E[Y|Y>Q p-1 ], so that τ p , >τ p+1 .
[0163] q p ∈(0,1),q p YesP p [Y|⊙ p ] is the qth quantile Q p ,
[0164] E[] is the expected value function.
[0165] Below about Fig.11 One embodiment of operation 1080 is described. After process 1000 is completed, process 1000 moves to end operation 1085.
[0166] Fig.11 is a process flow chart for estimating parameters of an unknown distribution. In some embodiments of the present disclosure, the unknown distribution is a distribution of events occurring on a monitored network. This unknown distribution is dynamically estimated by estimating the parameters of the distribution. Based on this estimation process, embodiments of the present disclosure are able to determine the probabilities of individual events occurring within the distribution. Then, in at least some embodiments, a brute force attack is determined based on the determined probabilities.
[0167] The unknown distribution is described by multiple parameters. Each parameter describes a characteristic of the unknown distribution. For example, in various embodiments, the multiple parameters describe two or more of the standard deviation, mean, and median of the distribution. In at least some embodiments, the distribution is modeled as a list of distributions P1[Y|⊙1], ..., Pm[Y|⊙ m ], where each parameter ⊙ p is a random process.
[0168] Below about Fig.11 Discussion process 1100. Fig.11 Determining updated distribution parameters includes selecting a domain for distribution as discussed above. In some embodiments, the process 1100 discussed below is derived from the above description of Fig. 9 The updater 851 discussed is executed.
[0169] In some embodiments, process 1100 is performed on multiple time series obtained through the disclosed embodiments. For example, with respect to the above discussion Figure 8 In some embodiments, process 1100 is performed to estimate distribution parameters for time series 802A and is performed separately to estimate distribution parameters for time series 802B. The set of time series is denoted as Y1, ...Y k Therefore, each Yi={y ij |j=1..n i} is a subset of the time series of the given data from the generating process P[Yi|⊙i], where each P[.|⊙i] can be a different family of probability distributions.
[0170] In some embodiments, the following Fig.11 One or more functions discussed are performed by hardware processing circuitry (e.g., 1202). In some embodiments, instructions (e.g., 1224) stored in memory (e.g., 1204, 1206) configure the hardware processing circuitry to perform one or more functions discussed below. In some embodiments, process 1100 discussed below is performed by a network management system or network security system (e.g., 110). In some embodiments, the following description of Fig.11 One or more functions discussed as part of operation 1080 are described above with respect to Fig.10 The discussed area is executed.
[0171] After starting operation 1102, process 1100 moves to operation 1104, which initializes the parameters of the distribution. In some embodiments, Gibbs sampling is used to estimate the parameters. For example, some embodiments calculate the conditional likelihood for each parameter under a stratified prior set for parameter selection. The parameters are then initialized based on the conditional likelihood.
[0172] In operation 1106, a parameter among the plurality of parameters is selected. In operation 1110, an additive parameter corresponding to the selected parameter is obtained via a link function. For example, in some embodiments, the additive parameter n t Obtained through:
[0173] η t :-g Φ [Φ t ]
[0174] in:
[0175] g Φ is the link function of the parameter Φ.
[0176] In operation 1120, the smoothing weights are sampled. In operation 1130, the additive parameter is updated based on the smoothing weights and the central moment corresponding to the parameter. For example, in some embodiments, the additive parameter η t Updated via:
[0177]
[0178] in:
[0179] is the smoothing weight,
[0180] y t+∈is the time series at time t+∈, and
[0181] M Φ is the central moment corresponding to the parameter Φ.
[0182] In operation 1140, the selected parameters are updated based on the reverse link function of the updated additive parameters. In some embodiments, the selected parameters are updated according to the following:
[0183] Φ t+∈ =g Φ -1 [η t+∈ ]
[0184] in:
[0185] Φ t+∈ is the updated parameter,
[0186] g Φ -1 is the inverse of the link function, and
[0187] η t+∈ is the additive parameter of the update.
[0188] Operation 1170 determines if there are additional parameters for the distribution. If so, processing returns to operation 1106. Otherwise, processing moves to end block 1185.
[0189] Notice, Fig.11 At least some embodiments of the process 1100 are directed to defining or otherwise updating a range for selecting a distribution in a multimodal model. For example, as discussed above, in some embodiments of the process 1100, the boundaries delineating the first parameter range and the second parameter range are adjusted based on the values of the time series. Then, the likelihood of a brute force attack is determined based on the adjusted boundaries. As discussed above with respect to Figure 8 As illustrated, adjusting the boundaries between the value ranges controls which distribution is compared to the time series at any particular time. The comparison of the time series to the distribution determines the likelihood that the values included in the time series were generated in the selected distribution. In some embodiments, this likelihood can be combined with other likelihoods of values from other time series (e.g., via Fisher's method).
[0190] Fig.12A block diagram of an example machine 1200 on which any one or more of the techniques (e.g., methods) discussed herein may be performed is illustrated. In alternative embodiments, the machine 1200 may operate as a standalone device or may be connected (e.g., networked) to other machines. In a networked deployment, the machine 1200 may operate in the capacity of a server machine, a client machine, or both in a server-client network environment. In an example, the machine 1200 may act as a peer machine in a peer-to-peer (P2P) (or other distributed) network environment. The machine 1200 may be a personal computer (PC), a tablet PC, a set-top box (STB), a personal digital assistant (PDA), a mobile phone, a smart phone, a network appliance, a network router, a switch or a bridge, a server computer, a database, a conference room device, or any machine capable of executing instructions (sequentially or otherwise) specifying actions to be taken by the machine. In various embodiments, the machine 1200 may perform the above with respect to the above Figures 1 to 11 Further, while only a single machine is illustrated, the term "machine" shall also be taken to include any collection of machines that individually or jointly execute a set (or multiple sets) of instructions to perform any one or more of the methodologies discussed herein, such as cloud computing, software as a service (SaaS), and other computer cluster configurations.
[0191] As described herein, examples may include or operate on logic or several components, modules, or mechanisms (hereinafter referred to as "modules"). A module is a tangible entity (e.g., hardware) that can perform a specified operation and can be configured or arranged in a certain manner. In one example, a circuit (e.g., internally or relative to an external entity such as other circuits) can be arranged as a module in a specified manner. In an example, all or part of one or more computer systems (e.g., stand-alone, client, or server computer systems) or one or more hardware processors can be configured as a module that operates to perform a specified operation by firmware or software (e.g., instructions, application parts, or applications). In an example, the software can reside on a machine-readable medium. In one example, the software causes the hardware to perform a specified operation when executed by the underlying hardware of the module.
[0192] Thus, the term "module" is understood to encompass a tangible entity, i.e., an entity that is physically constructed, specifically configured (e.g., hardwired), or temporarily (e.g., temporarily) configured (e.g., programmed) to operate in a specified manner or to perform part or all of any operations described herein. Consider an example in which a module is temporarily configured, without each module being instantiated at any time. For example, where a module includes a general-purpose hardware processor configured using software, the general-purpose hardware processor can be configured as various different modules at different times. The software can accordingly configure the hardware processor to, for example, constitute a particular module at one time instance and constitute different modules at different time instances.
[0193] The machine (e.g., computer system) 1200 may include a hardware processor 1202 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), a hardware processor core, or any combination thereof), a main memory 1204, and a static memory 1206, some or all of which may communicate with each other via an interconnection link (e.g., a bus) 1208. The machine 1200 may also include a display unit 1210, an alphanumeric input device 1212 (e.g., a keyboard), and a user interface (UI) navigation device 1214 (e.g., a mouse). In an example, the display unit 1210, the input device 1212, and the UI navigation device 1214 may be a touch screen display. The machine 1200 may additionally include a storage device (e.g., a drive unit) 1216, a signal generating device 1218 (e.g., a speaker), a network interface device 1220, and one or more sensors 1221 (such as a global positioning system (GPS) sensor, a compass, an accelerometer, or other sensor). The machine 1200 may include an output controller 1228, such as a serial (e.g., universal serial bus (USB), parallel, or other wired or wireless (e.g., infrared (IR), near field communication (NFC), etc.)) connection to communicate with or control one or more peripheral devices (e.g., a printer, a card reader, etc.).
[0194] The storage device 1216 may include a machine-readable medium 1222 on which is stored one or more sets of data structures or instructions 1224 (e.g., software) embodying or used by any one or more of the techniques or functionality described herein. The instructions 1224 may also reside, completely or at least partially, within the main memory 1204, within the static memory 1206, or within the hardware processor 1202 during execution thereof by the machine 1200. In an example, one or any combination of the hardware processor 1202, the main memory 1204, the static memory 1206, or the storage device 1216 may constitute a machine-readable medium.
[0195] Although the machine-readable medium 1222 is illustrated as a single medium, the term “machine-readable medium” may include a single medium or multiple media (eg, a centralized or distributed database and / or associated caches and servers) configured to store one or more instructions 1224 .
[0196] The term "machine-readable medium" may include any medium that is capable of storing, encoding, and / or carrying instructions for execution by the machine 1200 and causing the machine 1200 to perform any one or more of the techniques of the present disclosure, or capable of storing, encoding, or carrying data structures used by or associated with such instructions. Non-limiting examples of machine-readable media may include solid-state memory and optical and magnetic media. Specific examples of machine-readable media may include: non-volatile memory, such as semiconductor memory devices (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)) and flash memory devices; disks, such as internal hard disks and removable disks; magneto-optical disks; random access memory (RAM); solid-state drives (SSD); and CD-ROM and DVD-ROM disks. In some examples, the machine-readable medium may include non-transitory machine-readable media. In some examples, the machine-readable medium may include machine-readable media that is not a transient propagation signal.
[0197] The instructions 1224 may also be transmitted or received over a communication network 1226 using a transmission medium via the network interface device 1220. The machine 1200 may communicate with one or more other machines using any of a number of transmission protocols (e.g., frame relay, Internet Protocol (IP), transmission control protocol (TCP), user datagram protocol (UDP), hypertext transfer protocol (HTTP), etc.). Example communication networks may include a local area network (LAN), a wide area network (WAN), a packet data network (e.g., the Internet), a mobile telephone network (e.g., a cellular network), a plain old telephone (POTS) network, and a wireless data network (e.g., known as a wireless data network). The Institute of Electrical and Electronics Engineers (IEEE) 802.11 series of standards, known as IEEE802.16 series of standards), IEEE802.15.4 series of standards, long term evolution (LTE) series of standards, universal mobile telecommunication system (UMTS) series of standards, peer-to-peer (P2P) network, etc. In an example, the network interface device 1220 may include one or more physical jacks (e.g., Ethernet, coaxial, or telephone jacks) or one or more antennas to connect to the communication network 1226. In an example, the network interface device 1220 may include multiple antennas to use at least one of single input multiple output (SIMO), multiple input multiple output (MIMO), or multiple input single output (MISO) technology for wireless communication. In some examples, the network interface device 1220 may use multi-user MIMO technology for wireless communication.
[0198] Example 1 is a method performed by a hardware processing circuit system, including: obtaining a first time series of operating parameter values of a device attached to a network; comparing the values of the first time series with a first parameter value range and a second parameter value range; based on the comparison, determining that the values of the first time series are within the first parameter value range; based on the determination, selecting a first distribution from a plurality of distributions; comparing the first time series with the selected first distribution; based on the comparison, determining a first probability that a value in the first time series occurs in the selected first distribution; based on the first probability, determining a likelihood of a brute force attack on the network; based on the first time series, adjusting boundaries of the first parameter value range and the second parameter value range; and based on the adjusted boundaries, determining a second likelihood of a brute force attack.
[0199] In Example 2, the subject matter of Example 1 optionally includes performing a mitigating action based on the likelihood.
[0200] In Example 3, the subject matter of Example 2 may optionally include wherein the mitigating action comprises changing an access control policy of the network.
[0201] In Example 4, the subject matter of Example 3 optionally includes where the changing of the access control policy comprises programmatically configuring a firewall of the network.
[0202] In Example 5, any one or more of the subjects of Examples 1-4 optionally include where the adjustment of the boundary includes updating a threshold τ_p defining the boundary between the first domain and the second domain, τ_p being defined according to the following formula: τ_p:=E[Y|Y>Q_(p-1)], where: E[] is the expected value function, Y is the first time series, tp is the threshold between the first domain and the second domain, and Qp is the qth quantile of the negative binomial distribution.
[0203] In Example 6, the subject matter of any one or more of Examples 1-5 may optionally include wherein the adjusting of the boundary further comprises updating parameters defining the first distribution and the second distribution based on the first time series.
[0204] In Example 7, the subject matter of any one or more of Examples 1-6 may optionally include wherein the parameters are updated via exponential smoothing and a smoothing weight grid.
[0205] In Example 8, the subject matter of Example 7 may optionally include wherein at least one parameter of the distribution is updated according to: t+∈ =g Φ -1[η t +∈], where: Φ t+∈ is the updated parameter, g Φ is the link function for parameter Φ, Where: M_Φ is the central moment corresponding to parameter Φ, is the smoothing weight, ηt is g Φ [Φt], and y_(t+∈) is the sample value included in the first time series.
[0206] In Example 9, the subject matter of the undefined example optionally includes where the first distribution and the second distribution are included in a finite mixture of distributions.
[0207] In Example 10, any one or more of the subjects of Examples 1-9 optionally include determining a second time series of second operating parameter values for the device; selecting a second distribution based on the second time series; performing a second comparison of the second time series with the selected second distribution; based on the second comparison, determining a second probability that a second value in the second time series occurs in the selected second distribution; applying Fisher's method to the first probability and the second probability; and based on the application, generating a combined anomaly indicator, wherein the determination of the likelihood of a brute force attack is also based on the combined indicator.
[0208] In Example 11, the subject matter of Example 10 may optionally include where the first time series indicates a cumulative number of successful logins over time, and the second time series indicates a cumulative number of unsuccessful logins over time.
[0209] In Example 12, the subject matter of Example 11 optionally includes generating a third time series indicating user account information associated with each login indicated by the first time series and the second time series, and applying the Fisher method to the third time series for a second time, wherein generation of the combined indicator is also based on the second application.
[0210] In Example 13, any one or more of the subject matter of Examples 11-12 may optionally include wherein the logon is a Remote Desktop Protocol (RDP) logon attempt.
[0211] In Example 14, any one or more of the subjects of Examples 1-13 optionally include where the first time series indicates the time of day and day of week of failed logins and RDP connections, the timing of successful logins after failed attempts, login type information, reasons for login failures, a cumulative count of unique usernames that generated unsuccessful logins without subsequent successful logins, a count of failed login attempts, a count of RDP external IP addresses for each device, and a count of two devices having RDP inbound connections from a common device.
[0212] Example 15 is a non-transitory computer-readable medium, comprising instructions that, when executed, configure a hardware processing circuit to perform operations, the operations comprising: obtaining a first time series of operating parameter values of a device attached to a network; comparing the values of the first time series with a first parameter value range and a second parameter value range; based on the comparison, determining that the values of the first time series are within the first parameter value range; based on the determination, selecting a first distribution from a plurality of distributions; comparing the first time series with the selected first distribution; based on the comparison, determining a first probability that a value in the first time series occurs in the selected first distribution; based on the first probability, determining a likelihood of a brute force attack on the network; based on the first time series, adjusting boundaries of the first parameter value range and the second parameter value range; and based on the adjusted boundaries, determining a second likelihood of a brute force attack.
[0213] In Example 16, the subject matter of Example 15 may optionally include further comprising an operation of performing a mitigation action based on the likelihood.
[0214] In Example 17, the subject matter of Example 16 may optionally include wherein the mitigating action comprises changing an access control policy of the network.
[0215] In Example 18, the subject matter of Example 17 optionally includes wherein the changing of the access control policy comprises programmatically configuring a firewall of the network.
[0216] In Example 19, any one or more of the subjects of Examples 15-18 optionally include where the adjustment of the boundary includes updating a threshold τ_p defining the boundary between the first domain and the second domain, τ_p being defined according to the following formula: τ_p:=E[Y|Y>Q_(p-1)], where: E[] is the expected value function, Y is the first time series, tp is the threshold between the first domain and the second domain, and Qp is the qth quantile of the negative binomial distribution.
[0217] In Example 20, the subject matter of any one or more of Examples 15-19 may optionally include wherein the adjusting of the boundary further comprises updating parameters defining the first distribution and the second distribution based on the first time series.
[0218] In Example 21, the subject matter of any one or more of Examples 15-20 may optionally include wherein the parameters are updated via exponential smoothing and a smoothing weight grid.
[0219] In Example 22, the subject matter of Example 21 optionally includes wherein at least one parameter of the distribution is updated according to the following formula: Φt+∈=gΦ-1[ηt+∈], where: Φt+∈ is the updated parameter, gΦ is the link function for the parameter Φ, Where: M_Φ is the central moment corresponding to parameter Φ, is the smoothing weight, ηt is g Φ [Φt], and y_(t+∈) is the sample value included in the first time series.
[0220] In Example 23, the subject matter of the undefined example optionally includes where the first distribution and the second distribution are included in a finite mixture of distributions.
[0221] In Example 24, any one or more of the subjects of Examples 15-23 optionally include operations, which further include: determining a second time series of second operating parameter values for the device; selecting a second distribution based on the second time series; performing a second comparison of the second time series with the selected second distribution; based on the second comparison, determining a second probability that a second value in the second time series occurs in the selected second distribution; applying the Fisher method to the first probability and the second probability; and based on the application, generating a combined anomaly indicator, wherein the determination of the likelihood of a brute force attack is also based on the combined indicator.
[0222] In Example 25, the subject matter of Example 24 may optionally include where the first time series indicates a cumulative number of successful logins over time, and the second time series indicates a cumulative number of unsuccessful logins over time.
[0223] In Example 26, the subject matter of Example 25 optionally includes operations further comprising generating a third time series indicating user account information associated with each login indicated by the first time series and the second time series, and applying the Fisher method a second time to the third time series, wherein generation of the combined indicator is also based on the second application.
[0224] In Example 27, the subject matter of Example 26 optionally includes wherein the login is a Remote Desktop Protocol (RDP) login attempt.
[0225] In Example 28, any one or more of the subjects of Examples 15-27 optionally include where the first time series indicates one or more of the time of day and day of week of failed logins and RDP connections, the time of a successful login after a failed attempt, login type information, a reason for login failure, a cumulative count of unique usernames that generated unsuccessful logins without subsequent successful logins, a count of failed login attempts, a count of IP addresses external to RDP for each device, and a count of two devices having RDP inbound connections from a common device.
[0226] Example 29 is a system comprising: a hardware processing circuit system; one or more hardware memories storing instructions that, when executed, configure the hardware processing circuit to perform operations, the operations comprising: obtaining a first time series of operating parameter values of a device attached to a network; comparing the values of the first time series with a first parameter value range and a second parameter value range; based on the comparison, determining that the values of the first time series are within the first parameter value range; based on the determination, selecting a first distribution from a plurality of distributions; comparing the first time series with the selected first distribution; based on the comparison, determining a first probability that a value in the first time series occurs in the selected first distribution; based on the first probability, determining a likelihood of a brute force attack on the network; based on the first time series, adjusting boundaries of the first parameter value range and the second parameter value range; and based on the adjusted boundaries, determining a second likelihood of a brute force attack.
[0227] In Example 30, the subject matter of Example 29 may optionally include further comprising an operation of performing a mitigation action based on the likelihood.
[0228] In Example 31, the subject matter of Example 30 may optionally include wherein the mitigating action comprises changing an access control policy of the network.
[0229] In Example 32, the subject matter of Example 31 optionally includes where the changing of the access control policy comprises programmatically configuring a firewall of the network.
[0230] In Example 33, any one or more of the subjects of Examples 29-32 optionally include where the adjustment of the boundary includes updating a threshold τ_p defining the boundary between the first domain and the second domain, τ_p being defined according to the following formula: τ_p:=E[Y|Y>Q_(p-1)], where: E[] is the expected value function, Y is the first time series, tp is the threshold between the first domain and the second domain, and Qp is the qth quantile of the negative binomial distribution.
[0231] In Example 34, the subject matter of any one or more of Examples 29-33 may optionally include wherein the adjusting of the boundary further comprises updating parameters defining the first distribution and the second distribution based on the first time series.
[0232] In Example 35, the subject matter of any one or more of Examples 29-34 may optionally include wherein the parameters are updated via exponential smoothing and a smoothing weight grid.
[0233] In Example 36, the subject matter of Example 35 optionally includes wherein at least one parameter of the distribution is updated according to: Φt+∈=gΦ-1[ηt+∈], wherein: Φt+∈ is the updated parameter, gΦ is the link function for the parameter Φ, Where: M_Φ is the central moment corresponding to parameter Φ, is a smoothing weight, ηt is gΦ[Φt], and y_(t+∈) is a sample value included in the first time series.
[0234] In Example 37, the undefined subject matter of the example optionally includes where the first distribution and the second distribution are included in a finite mixture of distributions.
[0235] In Example 38, any one or more of the subjects of Examples 29-37 optionally include an operation, which further includes: determining a second time series of second operating parameter values for the device; selecting a second distribution based on the second time series; performing a second comparison of the second time series with the selected second distribution; based on the second comparison, determining a second probability that a second value in the second time series occurs in the selected second distribution; applying the Fisher method to the first probability and the second probability; and based on the application, generating a combined anomaly indicator, wherein the determination of the likelihood of a brute force attack is also based on the combined indicator.
[0236] In Example 39, the subject matter of Example 38 may optionally include where the first time series indicates a cumulative number of successful logins over time, and the second time series indicates a cumulative number of unsuccessful logins over time.
[0237] In Example 40, the subject matter of Example 39 optionally includes operations further comprising generating a third time series indicating user account information associated with each login indicated by the first time series and the second time series, and applying the Fisher method a second time to the third time series, wherein generation of the combined indicator is also based on the second application.
[0238] In Example 41, the subject matter of Example 40 optionally includes wherein the logon is a Remote Desktop Protocol (RDP) logon attempt.
[0239] In Example 42, any one or more of the subjects of Examples 29-41 optionally include where the first time series indicates one or more of the time of day and day of week of failed logins and RDP connections, the time of a successful login after a failed attempt, login type information, a reason for login failure, a cumulative count of unique usernames that generated unsuccessful logins without subsequent successful logins, a count of failed login attempts, a count of IP addresses external to RDP for each device, and a count of two devices having RDP inbound connections from a common device.
Claims
1. A method performed by a hardware processing circuit system, comprising: obtaining a first time series of values of operating parameters of a device attached to the network; comparing the operating parameter value of the first time series with a first parameter value range and a second parameter value range; Based on the comparison, determining that the operating parameter value of the first time series is within the first parameter value range; Based on the determination, selecting a first distribution from a plurality of distributions; comparing the first time series to the selected first distribution; determining, based on performing the comparison on the first time series, a first probability that a value of the operating parameter values in the first time series occurs within the selected first distribution; Determining a likelihood of a brute force attack on the network based on the first probability; Based on the first time series, adjusting a boundary between the first parameter value range and the second parameter value range to generate at least one of an updated first parameter value range or an updated second parameter value range; as well as A second likelihood of a brute force attack is determined based on the updated first parameter value range or the updated second parameter value range. The method of claim 1 , further comprising performing a mitigating action based on the likelihood. The method of claim 2 , wherein the mitigating action comprises changing an access control policy of the network. The method of claim 3 , wherein the changing of the access control policy comprises programmatically configuring a firewall of the network.
5. The method of claim 1 , wherein the adjustment of the boundary comprises updating a threshold value τ defining the boundary between the first parameter range and the second parameter range. p , τ p It is defined by the following formula: τ p : =E[Y|Y>Q p-1 ] in: E[] is the expected value function, Y is the first time series, τ p is the threshold of the boundary between the first parameter value range and the second parameter value range or the updated first parameter value range and the updated second parameter value range, and Q p is the qth quantile of the negative binomial distribution. The method of claim 1 , wherein the adjusting of the boundary further comprises updating parameters defining the first distribution. The method of claim 6 , wherein the parameters are updated via exponential smoothing and a smoothing weight grid.
8. The method according to claim 7, wherein at least one parameter of the first distribution is updated according to the following formula: Φ t+∈ =g Φ -1 [η t+∈ ], in: Φ t+∈ is the updated parameter, g Φ is the link function for parameter Φ, in: M Φ is the central moment corresponding to parameter Φ, is the smoothing weight, η t Yes Φ [Φ t ],as well as y t+∈ are the sample values included in the first time series.
9. The method according to claim 1, further comprising: determining a second time series of second operating parameter values for the device; selecting a second distribution based on the second time series; comparing the second time series to the selected second distribution; determining a second probability that a second value in the second time series occurs in the selected second distribution based on the comparing the second time series; applying Fisher's method to the first probability and the second probability; and Based on the application, a combined anomaly indicator is generated, wherein the determination of the likelihood of the brute force attack is also based on the combined indicator.
10. A system comprising: Hardware processing circuit system; one or more hardware memories storing instructions that, when executed, configure the hardware processing circuitry to perform operations comprising: obtaining a first time series of values of operating parameters of a device attached to the network; comparing the operating parameter value of the first time series with a first parameter value range and a second parameter value range; Based on the comparison, determining that the operating parameter value of the first time series is within the first parameter value range; Based on the determination, selecting a first distribution from a plurality of distributions; comparing the first time series to the selected first distribution; determining a first probability that a value of the operating parameter values in the first time series occurs within the selected first distribution based on the comparing the first time series; Determining a likelihood of a brute force attack on the network based on the first probability; Based on the first time series, adjusting a boundary between the first parameter value range and the second parameter value range to generate at least one of an updated first parameter value range or an updated second parameter value range; and A second likelihood of a brute force attack is determined based on the updated first parameter value range or the updated second parameter value range.
11. The system of claim 10, the operations further comprising performing a mitigating action based on the likelihood.
12. The system of claim 11, wherein the mitigating action comprises changing an access control policy of the network.
13. The system of claim 12, wherein the changing of the access control policy comprises programmatically configuring a firewall of the network.
14. The system of claim 10, wherein the adjustment of the boundary comprises updating a threshold τ defining the boundary between the first parameter range and the second parameter range. p , τ p It is defined by the following formula: τ p : =E[Y|Y>Q p-1 ] in: E[] is the expected value function, Y is the first time series, τ p is the threshold of the boundary between the first parameter value range and the second parameter value range or the updated first parameter value range and the updated second parameter value range, Q p is the qth quantile of the negative binomial distribution.
15. The system of claim 10, wherein the adjusting of the boundary further comprises updating parameters defining the first distribution.
16. The system of claim 15, wherein the parameters are updated via exponential smoothing and a smoothing weight grid.
17. The system of claim 16, wherein at least one parameter of the first distribution is updated according to the following formula: Φ t+∈ =g Φ -1 [η t+∈ ], in: Φ t+∈ is the updated parameter, g Φ is the link function for parameter Φ, in: M Φ is the central moment corresponding to parameter Φ, is the smoothing weight, η t is g Φ [Φ t y t+∈ are the sample values included in the first time series.
18. The system of claim 10, wherein the operation further comprises modeling the distribution of the first time series as a finite mixture of distributions P1[Y|⊙1], ...P m [Y|⊙ m ], where each parameter ⊙ p is a random process.
19. The system of claim 18, wherein the first distribution is included in a finite mixture of the distributions.
20. A non-transitory computer-readable storage medium comprising instructions that, when executed by a processor, configure a hardware processing circuit system to perform operations comprising: obtaining a first time series of values of operating parameters of a device attached to the network; comparing the operating parameter value of the first time series with a first parameter value range and a second parameter value range; Based on the comparison, determining that the operating parameter value of the first time series is within the first parameter value range; Based on the determination, selecting a first distribution from a plurality of distributions; comparing the first time series to the selected first distribution; determining, based on said comparing said first time series, a first probability that a value of said operating parameter values in said first time series occurs within said selected first distribution; Determining a likelihood of a brute force attack on the network based on the first probability; Based on the first time series, adjusting a boundary between the first parameter value range and the second parameter value range to generate at least one of an updated first parameter value range or an updated second parameter value range; as well as A second likelihood of a brute force attack is determined based on the updated first parameter value range or the updated second parameter value range.
