DETAILED DESCRIPTION OF EXEMPLARY EMBODIMENTS

[0038] This invention will be further understood in view of the following detailed description.

[0039] In accordance with the invention, a system and method for a violation of an email security policy of a computer system is disclosed herein. A violation of an email security policy can be defined in several ways. Such an email security policy may be explicit or implicit, and generally refers to any activity which may be harmful to the computer system. For example, an attachment to an email which contains a virus may be considered a violation of a security policy. Attachments which contain viruses can manifest themselves in several ways, for example, by propagating and retransmitting themselves. Another violation of a security policy may be the act of emailing attachments to addresses who do not have a need to receive such attachments in the ordinary course. Alternatively, the security policy may be violated by “spam” mail, which are typically unsolicited emails that are sent to a large number of email accounts, often by accessing an address book of a host email account. The method disclosed herein detects and tracks such security violations in order to contain them.

[0040] A model is defined which models the transmission of prior email through the computer system through the computer system. The model may be statistical model or a probabilistic model. The transmission of emails “through” the system refers to emails transmitted to email accounts in the system, email transmitted by email accounts in the system, and between email accounts within the system. The system accumulates statistics relating to various aspects of email traffic flow through the computer system. According to one embodiment, the model is derived from observing the behavior or features of attachments to emails. Another embodiment concerns modeling the behavior of a particular email account. Yet another embodiment models the behavior of the several email accounts on the system to detect “bad” profiles. The model is stored on a database, which may be either at a client or at a server, or at both locations.

[0041] The selected email transmission is typically chosen for some recent time period to compare with the prior transmission of email. Each email and/or its respective attachment is identified with a unique identifier so it may be tracked through the system. Various statistics relating to the emails are gathered. The probability that some aspect of the email transmission, e.g. an attachment, an email transmission, is violative of an email security policy is estimated by applying the model based on the statistics that have been gathered. Whether the email transmission is classified as violative of the email security policy is then transmitted to the other clients.

[0042] The system 10 , as illustrated in FIG. 1 , has two primary components, one or more clients 20 and one or more servers 40 . The client 20 is defined herein as a program integrated with an email server 22 , which monitors and logs email traffic 50 for one or more email accounts 26 , and which generates reports that are sent to the server 40 . The client 20 may run on a separate computer from the email server 22 , or on the same computer. The server 40 may run at a central location and receives reports from the client 20 in order to generate statistics and alerts about violations of email security policy which are distributed back to the clients 20 .

[0043] The client 20 also includes a database 24 , which stores information about all email attachments that pass through the mail server 22 to one or more email accounts 26 . (Transmission of the email to the respective account may be prevented if a violation of a security policy is detected.) The system 10 contains a component to integrate with the email sever 22 . In an exemplary embodiment, the client 20 is integrated with SENDMAIL using PROCMAIL. The client 20 also contains an analysis component 28 to compute the unique identifiers for attachments. The data analysis component 28 extracts statistics from the database 24 to report to the server 40 . A communication component 30 handles the communication between the client 20 and the server 40 .

[0044] When integrated with the mail server 22 , the client 20 processes all email. Each email is logged in the database 24 along with a set of properties associated with that email including a unique reference number for that email, the sending email account, the recipient email accounts, the number of recipients, the number of attachments, if any, the time and date of the email, the size in bytes of the email body, the size in bytes of the subject line, the number and list of “keywords” in the email subject line or body, other linguistic features of the email content (which may be a wide variety of features such as the number of nouns, or noun phrases, and/or the frequency distribution of words, or the frequency distribution of n-grams, or other such linguistic features commonly known in the state of the art), as well as other recorded properties of the email (some that may be inferred by application of a probabilistic, statistical or classification model which may label the email with some category of interest).

[0045] The mail server 22 extracts attachments from the email, if any, and computes a unique identifier for each attachment. The name of the attachment or the subject of the email is typically not sufficient information for tracking because one virus may be sent under several different names and subject lines since these fields are easily alterable by the malicious software. The system computes the MD5 hash of every binary attachment received to create the unique identifier, using the hexadecimal representation of the binary as input to the algorithm. (The MD5 is known in the art, and described in R. Rivest, “The MD5 Message Digest Algorithm,” Internet RFC 1321, Paril 1992, which is incorporated by reference in its entirety herein.) (Polymorphic viruses will have different identifiers for each instance of the virus.) A probabilistic model for the attachments may be created by training a Naive Bayes model on a training set of email attachments, described in U.S. patent Ser. application No. [not yet known], filed Jul. 30, 2002, entitled “System and Methods for Detection of New Malicious Executables,” which is incorporated by reference above.

[0046] This unique identifier is used to aggregate information about the same attachment propagated in different emails. This step if most effective if payload, e.g., the content of the email, such as the body, the subject, and/or the content of the attachment, is replicated without change during virus propagation among spreading emails and thus tracking the email attachments via this identifier is possible.

[0047] The client 20 stores a record containing the identifier and other information and statistics for each email and attachment in the database 24 . This information is typically transmitted to the server 40 , and such information is also transmitted from the server 40 to the client 20 for information that is received from other clients 20 , or where identifiers or models have been updated. By querying the database 24 with a list of the identifiers for known programs that are “malicious,” e.g., that violate the security policy, the administrator can determine the points of entry of emails having such programs as attachments into a network, and can maintain a list of the senders and recipients of these emails. Even if a logged attachment was not initially acknowledged as malicious but only later categorized to be so, since a record of all attachments is stored in the database the points of entry can still be recovered.

[0048] System 10 allows the system administrator to distinguish between email traffic containing non-malicious email attachments and email traffic containing malicious software attachments. Malicious programs that self-replicate will likely propagate at a significantly different rate than regular attachments sent within the environment in which the system 10 is installed. These differences may become more apparent as all email is monitored, and (temporal) statistics are gathered carefully within that environment to establish norms for email flows, as will be described below.

[0049] The system 10 uses the information stored in the database in several ways. Since the system 10 can determine the points of entry of a malicious attachment into a network, e.g., the recipient email account 26 and/or the client 20 associated with the email account 26 , this can greatly assist the cleanup associated with an email virus incident and can help the system administrator reduce and contain the associated damage.

[0050] In addition, the client 20 gathers statistics about the propagation of each malicious attachment through the site which is shared with the server 40 . The system may define an attachment as malicious or benign by extracting features of the attachment, and using a probabilistic model to determine whether the attachment is malicious or benign. A procedure for classifying attachments is described in U.S. patent application Ser. No. [not yet known], filed Jul. 30, 2002, entitled “System and Methods for Detection of New Malicious Executables,” which is incorporated by reference above.

[0051] The system also may define a probabilistic or statistical model relating to the behavior of attachments derived from these statistics or features. This allows a global view of the propagation of malicious attachments and allows the system 10 to quantify the threat of these attachments as described below. Some statistics that are reported for each malicious attachment is the prevalence of an attachment and the birth rate of an attachment. The prevalence is the number of occurrences an attachment was observed by the client 20 and the birth rate is the average number of copies of the attachment which are transmitted from the same email account 26 . Both of these statistics can be easily obtained from the database 24 .

[0052] Self-replicating viruses naturally have extremely high birth rates. If a client 20 detects an attachment with a very high birth rate, the client 20 can warn the server 40 that this attachment is a potential self replicating virus. The server 40 can in turn warn other clients 20 about this attachment which can reduce the spread of these types of viruses.

[0053] Many self-replicating viruses have a similar method of propagation, i.e., they transmit themselves to email addresses found on the address book of the host computer. This behavior may manifest itself in an extremely high birth rate for the attachment. While in some cases a large birthrate for an attachment would be normal, such as in a broadcast message, self-replicating viruses are characterized in that the message is transmitted from multiple email accounts 26 . In fact, the number of email accounts 26 that send the message depends on the number of email accounts 26 that open the attachment.

[0054] An exemplary method for detecting self-replicating viruses is to classify an attachment as self replicating if its birth rate is greater than some threshold t and the attachment is sent from at least l email accounts. If an email flow record is above the threshold t, the client 20 notifies the server 40 with the unique identifier of the attachment. The server 40 propagates the unique identifier to the clients 20 which instruct the mail server 24 to block all emails that contain an attachment with this unique identifier. In practice, these mails can be queued until a system administrator can determine whether or not they are malicious.

[0055] The server 40 runs at a central location and communicates with the clients 20 deployed at various mail servers 22 . The server 40 can typically be operated by a trusted third party and various networks can make agreements with this third party to provide the services described herein.

[0056] The server 40 has several functions. The server 40 may be responsible for propagating an updated list of unique identifiers associated with known malicious viruses to the clients 20 . This propagation is automated which allows for rapid update of the clients 20 immediately when a new malicious virus is discovered. The server 40 is responsible for aggregating statistics obtained from the reports from clients 20 which allows the system 10 to monitor violations of security policies at a global level. The information contained in each record is shown in FIGS. 2 - 3 , which illustrates screens of the user interface for system 10 . The fields correspond to information that the server 40 needs to either query the client 20 for more information, or to compute basic aggregate statistics.

[0057] Screen 200 ( FIG. 2 ) displays information concerning all emails which are transmitted through the system. For each email, a reference code 202 is assigned, the sender email account 204 , the recipient email account 206 , and the number of recipients 208 are noted. Also indicated is the number of attachments 210 , the size of the email 212 , and the time and date 214 of transmission. Finally, the email is classified as “interesting” or “not interesting” or a similar category, such as malicious, benign, or borderline, as will be described in greater detail below.

[0058] Screen 250 ( FIG. 3 ) illustrates a number of features that may be stored and displayed for each email. For example, further information on the sender 252 , e.g., sender's email, sender's name, etc., and information on the recipient 254 , e.g., recipient's email, recipient's name, etc., may be stored and displayed. However, it is also important in certain contexts to maintain the identify of email accounts in confidence. It is therefore important to have a de-identified user account which tracks a particular account, but which does not reveal the identity of the account. A privacy feature is accomplished in the exemplary embodiment by way of an MD5 hash algorithm, as described above, or equivalent which is applied to each email address, thereby creating a unique alphanumeric identifier 256 for the email, but which does not reveal the email address. Alternatively an alphanumeric code may be similarly created for the email address of the sender (not shown). The sender information 252 is blank in screen 250 . This may of de-identifying email may be a useful feature for a security personnel working with the system who may not have authorization to know the true email addresses that may cause alerts. In such instance, a higher authority may be required to inspect any such alerts and would have access to the mapping from the real email address to the unique identifier.

[0059] Information concerning attachments as illustrated in FIG. 4 . Screen 260 of the user interface of the exemplary embodiment illustrates that each attachment is represented by a unique MD5 hash identifier 262 , as discussed above. Information regarding the transmission of the attachment is stored and illustrated in table 264 . In particular, table 264 duplicates some of the information of screen 200 ( FIG. 2 ) and indicates the sender email account 266 , the recipient email account 268 , and the time and date of transmission 270 of each email which included the attachment. Further information recorded is the number of recipients 272 of the particular email that included the attachment, the total number of attachments 274 in that email, and the size of the attachment 276 . Further information is the level of “interest” 278 of the attachment, which is a numerical figure generated, for example, by a probabilistic model such as Naive Bayes, regarding whether the attachment is malicious, benign or borderline, as determine by a virus scanner, or by the technique described in U.S. patent application Ser. No. [not yet known], filed Jul. 30, 2002, entitled “System and Methods for Detection of New Malicious Executables,” which is incorporated by reference above. Table 280 includes the classification malicious, benign or borderline, which is derived from the level of interest 278 , above. Additional information about the birthrate, and other statistics about the attachment are recorded and displayed in screen 260 .

[0060] This information may be stored on database 24 of client 20 and distributed to the server 40 (and database 42 ), and in turn to others clients 20 , which could update its local database 24 by including the unique attachment identifier along with its classification as malicious, so that any future emails that appear with an attachment whose MD5 hash matches the unique identifier would cause each client to alert on that email as containing a malicious attachment. MySQL, for example, may be used in the exemplary embodiment, which is a well-known open source database system.

[0061] The server 40 also contains a data analysis component 44 which performs the analysis over these records, such as computation or updating of statistics in the database 42 about attachments or emails, as well as application of probabilistic or statistical models or tests in order to generate alerts of emails or attachments that violate security policy. For example, a model which is used to classify an attachment as benign, malicious, or borderline may be performed at the data analysis component 44 . This model may be updated with additional training data, which may be different from the model that is used to classify attachments at the client 20 . A communication component 46 manages the communication with multiple clients 20 . The communication between the server 40 and the client 20 consists of messages passed on a secured channel using encryption and authentication mechanisms.

[0062] When a client 20 reports an incident of a received email attachment that is violative of a security policy, it may report a unique incident identification number, the unique identifier of the attachment, the date and time of the attack, the prevalence, and the birth rate.

[0063] Additional statistics may be computed for each attachment and stored on databases 24 / 42 and displayed, for example, in table 280 of screen 260 of the user interface. A virus incident is the fraction of the total number of clients 20 within an organization infected by a particular virus, due to a single initial infection from outside the organization. Since each attachment is saved in the local database 24 with a Unique identifier and malicious or benign classification, this value is simply the number of times each malicious unique identifier appears in the local database 24 . The lifespan is the length of time a virus is active. This value is calculated by subtracting the first time a virus is seen from its last occurrence in the local repository. This values reports the amount of time a virus was free to cause damage to a network before it was detected. The Incident rate is the rate at which virus incidents occur in a given population per unit time, normalized to the number of clients 20 in the population. This is calculated by the server 40 based on the virus incident values reported by the local server. The death rate is the rate at which a virus is detected. This is calculated by the server 40 by taking the average lifespan of the virus. The system prevalence is a measure at the system level of the total number of clients 20 infected by a particular virus. This value is calculated by the central repository by summing over the number of local hosts reporting the same virus. The threat is the measure of how much of a possible danger a virus may be. In an exemplary embodiment, threat is calculated as the incident rate of a virus added to the prevalence of a virus divided by the total number of participating clients 20 and the total number of viruses. Spread is a measure of the global birth rate of a virus. This is calculated by taking the average of the birth rates reported by the participating clients 20 . These metrics may be directly implemented by computing SQL aggregates over the databases (both local 24 and central 42 ). Each time a client 20 determines that an attachment is a virus, it sends a report to the server 40 , and the server 40 updates it statistics for that virus.

[0064] The system 10 may also gather statistics about the behavior and features of individual email accounts 26 , which is a representation of the users of these accounts. The information gathered about individual emails, as well as email accounts themselves, is useful to detecting violations of an email security policy. For example, email account statistics may be derived for recipient and sender email addresses recorded in the database. The statistics gathered about the prior transmission of email to and from a particular email account can be used as training data to create a probabilistic or statistical model of an email account. This model provides a profile of the past or baseline behavior patterns of a particular email account. The selected behavior may refer to a particular time frame of interest, e.g., the previous month. Where the selected behavior of the particular email account deviates from this profile of prior or baseline behavior, the system 10 may issue an alert that a violation of an email security policy has occurred.

[0065] This profile of behavior patterns may be represented as a histogram, for example. A histogram is a way of graphically showing the characteristics of the distribution of items in a given population of samples. In the exemplary embodiment, histograms are used to model the behavior of particular email accounts. From a training set, e.g., the statistics as discussed above, a histogram is constructed to represent the baseline behavior of an email account. A histogram is also created to represent selected behavior of the email account.

[0066] Histograms may model statistics, e.g., events or operations, which are accumulated over a fixed time period. Each bin in the histogram counts some number of events in fixed time periods. For example, a histogram may record the average number of emails sent by an email account each day during the previous month, wherein each bin represents a day, hour, or other time period. Alternatively, histograms may model statistics accumulated irrespective of a time period. In such case, each bin is not a fixed time period, but some other feature. For example, over a set of emails from an arbitrary time period (gathered over a month, or gathered over a year, etc.) a histogram recording the number of email sent to a distinct recipient, wherein each bin represents a recipient, for example.

[0067] FIG. 5 illustrates a screen 300 in the user interface of the exemplary embodiment, which illustrates histograms that may be stored for an email account 302 In the example, statistics are gathered for an email account 302 over a predetermined period of time, e.g., the previous twelve months. The system counts the number of emails sent by this email account 302 to a specific recipient. Table 304 shows each recipient email address 306 and the relative frequency 308 at which user account 302 has emailed each recipient. In histogram 310 , each recipient would be considered a bin 312 , which indicates the frequency of emails 314 for each recipient. If an email account has sent emails over the past twelve months to 900 different email accounts, for example, then the email account's profile histogram would have 900 bins. A histogram computed over the twelve months would serve as a statistical model of baseline behavior of the email account. The histogram's bins can be ordered from “most frequent” recipient to “least frequent” recipient and display these as a bar graph 310 (as in FIG. 5 ), or alternatively, the statistics may be represented as a continuous function or a plotted graph. The bins of the histogram may be ordered differently, by for example, sorting the recipient names, or grouping recipients according to email domain. A histogram of selected behavior may include bins for each email recipient, and taken over the selected time period.

[0068] A sequential profile can be represented which is irrespective of the quanta of time measured (non-stationary), but which instead uses each email as a measurement point. With continued reference to FIG. 5 , plot 320 illustrates the number of recipients 322 who received email from user account 302 . The list grows over the history of recorded emails as more emails 324 are sent. Graph 320 monotonically increases for each sequential email measured. The growth rate of this plot indicates a profile of the email account. A plot that is very slowly increasing indicates that the email account does not exchange emails with very many new email accounts. While another email account may have a very fast growing profile, perhaps indicating that the user of the email account may be contacted by very many new people. A histogram for normal behavior may be taken over one time period, and histogram for new behavior may be taken over a second time period. Graph 330 illustrates the distinct number of recipient per 50 emails sent (dashed line 332 ) and the distinct number of recipients per 20 emails sent (dotted line 334 ). As another example, the first 100 emails sent in order over some time period by an email account were sent to ten distinct email addresses. In the 101 ^st -110 ^th emails, no new email addresses are seen that are distinct from those seen in the first 100 emails. However, two new distinct email addresses are seen in the 112 ^th email. For this email, we have a net gain of two more emails. Such growth rates are statistics that may be used to detect violations of security policy.

[0069] Once such histograms have been created, the histogram of the baseline behavior is compared with the histogram of the selected behavior to determine whether the new behavior represents a deviation that may be classified as a violation of email security policy. There are many known methods to compute the histogram dissimilarity. Generally such methods may be divided into two categories: One method is using a histogram distance function; the other method is to use a statistics test. A histogram can be represented by a vector.

[0070] Histograms may be compared with the L 1 form distance equation. Histogram intersection is represented in equation (1), where X and Y are vectors representing the normal behavior histogram and the new behavior histogram. M is the number of bins in histogram. 1 $\begin{array}{cc}L\left(X,Y\right)=1-\frac{\sum _{i=0}^{M-1}\min \left(X\left[i\right],Y\left[i\right]\right)}{\min \left(\sum _{i=0}^{M-1}X\left[i\right],\sum _{i=0}^{M-1}Y\left[i\right]\right)}& \left(1\right)\end{array}$

[0071] When the sums of X[i] and Y[i] are equal, the histogram intersection formula of equation (1) may be simplified to the L 1 form distance equation (2): 2 $\begin{array}{cc}{L}_1\left(X,Y\right)=\sum _{i=0}^{M-1}X\left[i\right]-Y\left[i\right]& \left(2\right)\end{array}$

[0072] Alternatively, histograms may be compared with the L 2 form distance equation (3): 3 $\begin{array}{cc}{L}_2\left(X,Y\right)=\sum _{i=0}^{M-1}{\left(X\left[i\right]-Y\left[i\right]\right)}^2& \left(3\right)\end{array}$

[0073] The L 1 and L 2 form equations assume that the individual components of the feature vectors, e.g., the bins of the histograms, are independent from each other. Each of the bins are taken to contribute equally to the distance, and the difference of content between the various bins is ignored.

[0074] Other distance equations are the weighted histogram difference equations, e.g., the histogram quadratic distance equation and the histogram Mahalanobis distance equation. The histogram quadratic difference equation (4) considers the difference between different bins.

D ( X, Y )=( X−Y ) ^T A ( X−Y ) (4)

[0075] In equation (4), A is a matrix and a _ij denotes the similarity between elements with index i and j. A symmetry is assumed, such that a _ij =a _jl , and a _ij =1.

[0076] The Mahalanobis distance is a special case of the quadratic distance equation. The matrix A is given by the covariance matrix obtained from a set of training histograms. Here, the elements in the histogram vectors are treated as random variables, i.e., X=[x _{0, x 1} , . . . , x _M−1 ]. The covariance matrix B is defined as b _lj =Cov(x _l , x _l ). The matrix A is thus defined as A=B ⁻¹ . When the x _l are statistically independent, but have unequal variance, matrix B is a diagonal matrix: 4 $\begin{array}{cc}B=\left[\begin{array}{c}{\sigma }_0^2,0,0,\dots ,0\\ 0,{\sigma }_0^2,0,\dots ,0\\ 0,\dots ,0,0\\ 0,\dots 0,0,{\sigma }_{M-1}^2\end{array}\right]& \left(5\right)\end{array}$

[0077] This method requires a sufficiently large training set (of prior email transmission statistics) in order to allow the covariance matrix to accurately represent the training data.

[0078] The chi-square test is used to test if a sample of data came from a population with a specific distribution. It can be applied to any uni-variance distribution for which it is possible to calculate the cumulative distribution function. However, the value of chi-square test statistic depends on how the data is binned, and it requires a sufficient sample size. The chi-square test is represented by equation (6): 5 $\begin{array}{cc}{\chi }^2=\sum _{i=1}^k{\left(O_i-E_i\right)}^2/E_i& \left(6\right)\end{array}$

[0079] where k is the number of bins O _l is the observed frequency for bin i, and E _l is the expected frequency. The expected frequency is calculated as:

E _l =N ( F ( Y _u )− F ( Y _l )). (7)

[0080] where F is the cumulative distribution function, Y _u is the upper limit for class i, Y _l is the lower limit for class i, and N is the sample size.

[0081] The Kolmogorov-Simironov test (the “KS test”) is a statistical test which is designed to test the hypothesis that a given data set could have been drawn from a given distribution, i.e., that the new behavior could have been drawn from the normal behavior. The KS test is primarily intended for use with data having a continuous distribution, and with data that is independent of arbitrary computational choice, such as bin width. The result D is equal to the maximum difference between the cumulative distribution of data points.

D =max{| F ′( x )− F ( x )|}, F ′( x )=(num_of_samples≦ x )/ N (8)

[0082] and where N is total number of samples The KS test does not depend on the underlying cumulative distribution function which is being tested, and it is an exact test (when compared with the Chi-Square test, which depends on an adequate sample size for the approximations to be valid). The KS test may only be applied to continuous distribution; it tends to be more sensitive near of the center of the distribution than at the tails.

[0083] The modeling of the behavior of an email account may include defining a model based on the time of day in which emails are transmitted by a particular email account. FIG. 6 illustrates screen 400 , which compares such email transmission for user account 402 . Histogram 404 illustrates the average number of emails 406 sent for each bin 408 , which represents each hour of the 24 hours in a day. The data in histogram 404 is accumulated for a predetermined period of time, e.g., the entire period that user account 402 has been tracked by the system 10 (time period 410 ). Histogram 412 is created for email transmission during a selected period of time being analyzed, e.g., the last month (time period 414 ). Histogram 412 illustrates the average number of emails 416 sent during each hour as represented by bins 418 . The histogram 404 of baseline behavior is compared with the histogram 412 of the selected behavior, with a comparison equation such as the Mahalanobis distance equation, above, to produce a distance result 320 . A threshold is set, which determines whether such a calculated difference is normal or may possibly violate security policy. The threshold may be determined by training on known data representative of email account behavior which violated security policy, when compared with known, normal, email behavior. The histogram 404 of the baseline behavior of user email account 302 shows that emails are rarely sent early in the morning. Thus, a violation in the security policy may be detected if a series of email are transmitted from user email account 302 at such time of day. Similarly, the modeling of the behavior of an email account may include defining a model based on the size of the emails that are transmitted by an email account or on the number of attachments that are transmitted by the email account

[0084] Another method for defining a model relating to the transmission of emails from one of the email accounts is based on the email addresses of the recipients of emails transmitted by the particular email account. Thus, another statistic or feature gathered by the method in accordance with the invention is the email addresses of recipients in each email. The recipients of the emails may be grouped into “cliques” corresponding to email addresses historically occurring in the same email.

[0085] A clique is defined as a cluster of strongly related objects in a set of objects. A clique can be represented as a subset of a graph, where nodes in the graph represent the “objects” and arcs or edges between nodes represent the “relationships” between the objects. Further, a clique is a subset of nodes where each pair of nodes in the clique share the relationship but other nodes in the graph do not. There may be many cliques in any graph.

[0086] In this context, the nodes are email addresses (or accounts) and the edges represent the “emails” (and or the quantity of emails) exchanged between the objects (email accounts). Each email account is regarded as a node, and the relationship between them is determined by the to:, from:, and cc: fields of the emails exchanged between the email accounts. As illustrated in FIG. 7, a selected email account 100 induces its own set of cliques 110 a , 110 b , 110 c , which are clusters of email accounts 120 of which it is a member. Each member in the clique has been determined to historically exchange emails 130 with each other. This modeling of email cliques is based on the premise that a user's “social cliques” and the nature of the relationship between members of a clique can be revealed by their “email cliques.”

[0087] The relationship between nodes that induces the cliques can be defined under different periods of time, and with different numbers of emails being exchanged, or other features or properties. For example, an edge (as represented by line 130 in FIG. 7 ) between email account UserA@z.com and email account UserB@z.com may be represented if UserA and UserB have exchanged at least N emails over the time period T. (As one varies N, the cliques revealed may change.) As another example, an edge between UserC and UserD may be represented if they have exchanged at least N emails with each other in the time period T, and each email is at least K bytes long. Such features of emails are based upon the kind of information an analyst may wish to extract from a set of emails. As a further example, one may define the clique relationship to be the set of accounts that exchange at least N emails per time period T and which include certain string of text S. (Further details concerning clique finding algorithms and related problems are disclosed in Cliques, Coloring and Satisfiability: Second Dimacs Implementation Challenge, D. Johnson and M. Trick, Ed., 1993, which is incorporated by reference in its entirety herein.)

[0088] FIG. 7 illustrates the email behavior of the user of email account 100 . For example, the three clusters may represent cliques of social acquaintances 110 a , clients 110 b , and coworkers 110 c . (Although four email accounts are shown in each clique 110 a , 110 b , and 110 c , it is understood that the number of email accounts may be larger or smaller depending upon the historical email use of the particular email accounts.) Each of these groups of users with their own email accounts 120 , have a relationship with the user of email account 100 . Members of different cliques, i.e., social acquaintances 110 a and clients 110 b are unlikely to have common interests or concerns. Thus, it is unlikely that the user of email account 100 would send the same email to both cliques. More particularly, it is unlikely that email account 100 would send an email 140 addressed to both an email account in clique 110 a and an email account in clique 110 b (illustrated in dotted line).

[0089] Cliques are determined according to any number of known methods. In the exemplary embodiment, cliques are modeled as described in C. Bron and J. Kerbosch. “Algorithm 457: Finding All Cliques of an Undirected Graph,” Communications of ACM, 16:575-577, 1973, which is incorporated in The Appendix and the attached routine Clique_finder.

[0090] First, the graph is built by selecting all of the rows from the email table in the database. As illustrated in FIG. 2 , above each row contains the sender 204 , and the recipient 206 . The subject line may also be stored (although not illustrated in FIG. 2 ).

[0091] A first step is to check an aliases file against the sender and recipient to map all aliases to a common name. For instance, a single user may have several accounts. This information, if available, would be stored in an aliases file.

[0092] The edge between sender and recipient is updated (or added if it doesn't already exist). (The edge is represented as line 130 in FIG. 7 .) Each edge of the graph may have associated with it ( 1) the number of emails that traversed that edge and (2) a weighted set of subject words where each word has a count of the number of times it occurred. The edge's weight is incremented by one, and the weighted set of subject words associated with the edge is augmented by the set of subject words from the current message. Cliques are represented in screen 500 of the user interface in FIG. 8 . Cliques 502 , 504 , and 506 are displayed, along with the most common subject words in emails transmitted among members of the clique.

[0093] A next step is pruning the graph. The user inputs a minimum edge weight, or minimum number of emails that must pass between the two accounts to constitute an edge, and any edges that don't meet that weight are eliminated. For example, the minimum number of emails may be determined from the average number of emails sent by the email account over a similar time period.

[0094] Subsequently, the cliques are determined. Throughout this process, there exist four sets of data: (1) * compsub* represents a stack of email user accounts representing the clique being evaluated. Every account in * compsub* is connected to every other account. (2) * candi da tes* represents a set of email user accounts whose status is yet to be determined. (3) *not* represents a set of accounts that have earlier served as an extension of the present configuration of * compsub* and are now explicitly excluded. (4) * cliques* represents a set of completed cliques

[0095] In the exemplary embodiment, these are implemented using the Java Stack and HashSet classes rather than the array structure suggested in the Bron & Kerbosch in The Appendix and the routine Clique_finder attached herein.

[0096] The algorithm is a recursive call to extendClique(). There are five steps in the algorithm: Step 1 is the selection of a candidate, i.e., an email user account which may be prospectively added to the clique. Step 2 involves adding the selected candidate to *compsub*. Step 3 creates new sets *candidates* and *not* from the old sets by removing all points not connected to the selected candidate (to remain consistent with the definition), keeping the old sets intact. Step 4 is calling the extension operator to operate on the sets just formed. The duty of the extension operator is generate all extensions of the given configuration of * compsub* that it can make with the given set of candidates and that do not contain any of the points in *not*. Upon return, step 5 is the removal of the selected candidate from * compsub* and its addition to the old set *not*.

[0097] When *candidates* and *not* are both empty, a copy of *compsub* is added to * cliques*. (If *not* is non-empty it means that the clique in * compsub* is not maximal and was contained in an earlier clique.) A clique's most frequent subject words are computed by merging and sorting the weighted sets of subject words on each edge in the clique.

[0098] If we reach a point where there is a point in *not* connected to all the points in * candidates*, the clique determination is completed (as discussed in The Appendix). This state is reached as quickly as possible by fixing a point in *not* that has the most connections to points in *candidates* and always choosing a candidate that is not connected to that fixed point.

[0099] A clique violation occurs if a user email account sends email to recipients which are in different cliques. If an email 140 is detected, this occurrence of an email having a recipient in two different cliques may be considered a clique violation, and may indicate that either a) email account 100 made a mistake by sending an inappropriate message to either a social acquaintance or to a client or b) a self-replicating email attachment has accessed the address book for the email account 100 and is transmitting itself to email accounts in the address-book without knowledge the cliques 110 a , 110 b , 110 c of email account 100 .

[0100] A strength of the clique violation may be measured by counting the number of such violations in a single email, e.g., the number of recipients who are not themselves part of the same clique, and/or the number of emails being sent, or other features that may be defined (as the system designer's choice) to quantify the severity of the clique violation. (For example, if email account 100 sent one message to 15 recipients, and one of these recipients is not a member of a clique that the other 14 belong to, that may be considered a minor violation compared with another email that is directed to 15 recipients none of whom are members of the same clique.) The strength of the violation may be used to set conditions (or thresholds) which are used to provide alerts in the system 10 . Alerts may then be generated based upon the strength of the violation. In another embodiment, those recipients that receive few emails from the sender may be weighted higher than those recipients that receive many emails from the sender.

[0101] Clique violations may also be determined from multiple email messages, rather than from just one email. For example, if a set of emails are sent over some period of time, and each of these emails are “similar” in some way, the set of email accounts contained in those emails can be subjected to clique violation tests. Thus, the email recipients of email sent by a particular use is used as training data to train a model of the email account.

[0102] If a specific email account is being protected by this method of modeling cliques and detecting clique violations, such violations could represent a misuse of the email account in question. For example, this event may represent a security violation if the VP of engineering sends an email to the CEO concurrently with a friend who is not an employee of the VP's company. Similarly, a clique violation would occur when a navy lieutenant sends a secret document to his commanding officer, with his wife's email account in the CC field. These are clique violations that would trigger an alert.

[0103] The techniques described herein can also be used a) to detect spam emails (which may or may not and generally do not have attachments, and b) to detect spammers themselves. Spam generally has no attachments, so other statistics about email content and email account behavior are needed to be gathered here by system 10 in order to also detect spam. Spam can be detected by considering clique violations. In particular, if an email account sends or receives emails from other email accounts that are not in the same clique, an alert may be issued which would indicate that such email transmissions are likely spam.

[0104] The methods described above generally refer to defining probabilistic or statistical models which define the behavior of individual email accounts. Also useful are models relating to statistics for emails transmitted by the plurality of email accounts on the computer system.

[0105] Detecting email accounts that are being used by spammers may allow an internet service provider or server 40 to stop spam from spreading from their service by shutting down an email account that has been detected as a generator of spam. To detect spammers, these email accounts would have a certain profile of email use that may be regarded as a bad profile as determined by supervised machine learning process, for example. Thus, the notion of profiling i.e., gathering statistics about an email account's behavior, is used here as well. According to this embodiment, email profiles are compared to other email profiles, rather than comparing statistics about emails to profiles.

[0106] Individual profiles may be represented by histograms in screen 550 of the user interface as illustrated in FIG. 9 for user 552 . Histogram 554 indicates the average number of emails sent on particular days of the week 556 , and sorted in bins for daytime 558 , evening 560 , and night 562 . Similarly, histogram 564 indicates the average size (in bytes) of emails sent on particular days of the week 566 , and sorted in bins for daytime 568 , evening 570 , and night 572 . Histogram 574 indicates the average number of recipients for each email sent on particular days of the week 576 , and sorted in bins for daytime 578 , evening 580 , and night 582 .

EXAMPLE

[0107] Detection of a “spammer” may be performed by comparing email account profiles, such as those illustrated in FIG. 9 . The following three profiles, or models, are created from statistics gathered by the system:

[0108] Profile 1 : Histogram of average number of emails sent per minute and per day by a user account computed over a one week period. (Table 1) 1

[0109] Profile 2 : Histogram of average number of recipients per email for morning, day, night. (Table 2) 2

[0110] Profile 3 : Histogram of cumulative number of distinct email account recipients per email sent (which may be plotted as a function, or even represented by a closed form functional description modeled as a linear function, or a quadratic function, etc.) 3

[0111] Given these three profiles, Account A appears to have a profile showing very modest use of emails, with few recipients. Account B on the other hand appears to be a heavy transmitter of emails. In addition, there seems to be evidence that the behavior of Account B is indicative of a ‘drone’ spammer. Such determination may be made by comparing the histograms of Account A (considered a “normal” user) with the histograms of Account B, and determining the difference between the two. Equations (1)-(8), above, are useful for this purpose. For example, the histogram of Table 2 indicates that the behavior of Account B may be consistent with running a program that is automaticaaly sending emails to a fixed number of recipients (e.g., 15 ), and the histogram of Table 3 indicates that there is a very large number of email addresses in Account B's address book. In the illustration, Account B has already generated 1236 distinct address by email 55 . The inference can therefore be made that Account B is a spammer. This type of profile can be used to find other similar profiles of other accounts indicative of other spammers.

[0112] It will be understood that the foregoing is only illustrative of the principles of the invention, and that various modifications can be made by those skilled in the art without departing from the scope and spirit of the invention.