Patent Application
20110208861
The present invention relates to computer networks, and in particular, to a object classification.
Computer networks and systems have become indispensable tools for modern business. Modern enterprises use such networks for communications and for storage. The information and data stored on the network of a business enterprise is often a highly valuable asset. Modern enterprises use numerous tools to keep outsiders, intruders, and unauthorized personnel from accessing valuable information stored on the network. These tools include firewalls, intrusion detection systems, and packet sniffer devices. However, once an intruder has gained access to sensitive content, there is no network device that can prevent the electronic transmission of the content from the network to outside the network. Similarly, there is no network device that can analyse the data leaving the network to monitor for policy violations, and make it possible to track down information leeks. What is needed is a comprehensive system to capture, store, and analyse all data communicated using the enterprises network.
Objects can be extracted from data flows captured by a capture device. Each captured object can then be classified according to content. In one embodiment, the present invention includes determining whether a captured object is binary or textual in nature, and classifying the captured object as one of a plurality of textual content types based tokens found in the captured object if the captured object is determined to be textual in nature.
The present invention is illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings in which like reference numerals refer to similar elements and in which:
Although the present system will be discussed with reference to various illustrated examples, these examples should not be read to limit the broader spirit and scope of the present invention. Some portions of the detailed description that follows are presented in terms of algorithms and symbolic representations of operations on data within a computer memory. These algorithmic descriptions and representations are the means used by those skilled in the computer science arts to most effectively convey the substance of their work to others skilled in the art. An algorithm is here, and generally, conceived to be a self-consistent sequence of steps leading to a desired result. The steps are those requiring physical manipulations of physical quantities. Usually, though not necessarily, these quantities take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared and otherwise manipulated.
It has proven convenient at times, principally for reasons of common usage, to refer to these signals as bits, values, elements, symbols, characters, terms, numbers or the like. It should be borne in mind, however, that all of these and similar terms are to be associated with the appropriate physical quantities and are merely convenient labels applied to these quantities. Unless specifically stated otherwise, it will be appreciated that throughout the description of the present invention, use of terms such as “processing”, “computing”, “calculating”, “determining”, “displaying” or the like, refer to the action and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantities within the computer system memories or registers or other such information storage, transmission or display devices.
As indicated above, one embodiment of the present invention is instantiated in computer software, that is, computer readable instructions, which, when executed by one or more computer processors/systems, instruct the processors/systems to perform the designated actions. Such computer software may be resident in one or more computer readable media, such as hard drives, CD-ROMs, DVD-ROMs, read-only memory, read-write memory and so on. Such software may be distributed on one or more of these media, or may be made available for download across one or more computer networks (e.g., the Internet). Regardless of the format, the computer programming, rendering and processing techniques discussed herein are simply examples of the types of programming, rendering and processing techniques that may be used to implement aspects of the present invention. These examples should in no way limit the present invention, which is best understood with reference to the claims that follow this description.
Networks
In
One embodiment of the present invention is now illustrated with reference to
There are various other possible configurations. For example, the router 12 can also forward a copy of all incoming data to the capture system 22 as well. Furthermore, the capture system 22 can be configured sequentially in front of, or behind the router 20, however this makes the capture system 22 a critical component in connecting to the Internet 12. In systems where a router 12 is not used at all, the capture system can be interposed directly between the LAN 10 and the Internet 12. In one embodiment, the capture system 22 has a user interface accessible from a LAN-attached device, such as a client 16.
In one embodiment, the capture system 22 intercepts all data leaving the network. In other embodiments, the capture system can also intercept all data being communicated inside the network 10. In one embodiment, the capture system 22 reconstructs the documents leaving the network 10, and stores them in a searchable fashion. The capture system 22 can then be used to search and sort through all documents that have left the network 10. There are many reasons such documents may be of interest, including network security reasons, intellectual property concerns, corporate governance regulations, and other corporate policy concerns.
Capture System
One embodiment of the present invention is now described with reference to
The captured raw data is then passed to a packet capture module 26. In one embodiment, the packet capture module 26 extracts data packets from the data stream received from the network interface module 24. In one embodiment, the packet capture module 26 reconstructs Ethernet packets from multiple sources to multiple destinations for the raw data stream.
In one embodiment, the packets are then provided the object assembly module 28. The object assembly module 28 reconstructs the objects being transmitted by the packets. For example, when a document is transmitted, e.g. as an email attachment, it is broken down into packets according to various data transfer protocols such as Transmission Control Protocol/Internet Protocol (TCP/IP) and Ethernet. The object assembly module 28 can reconstruct the document from the captured packets.
One embodiment of the object assembly module 28 is now described in more detail with reference to
In one embodiment, the reassembler 36 begins a new flow upon the observation of a starting packet defined by the data transfer protocol. For a TCP/IP embodiment, the starting packet is generally referred to as the “SYN” packet. The flow can terminate upon observation of a finishing packet, e.g., a “Reset” or “FIN” packet in TCP/IP. If now finishing packet is observed by the reassembler 36 within some time constraint, it can terminate the flow via a timeout mechanism. In an embodiment using the TPC protocol, a TCP flow contains an ordered sequence of packets that can be assembled into a contiguous data stream by the reassembler 36. Thus, in one embodiment, a flow is an ordered data stream of a single communication between a source and a destination.
The flown assembled by the reassember 36 can then be provided to a protocol demultiplexer (demux) 38. In one embodiment, the protocol demux 38 sorts assembled flows using the TCP Ports. This can include performing a speculative classification of the flow contents based on the association of well-known port numbers with specified protocols. For example, Web Hyper Text Transfer Protocol (HTTP) packets—i.e., Web traffic—are typically associated with port 80, File Transfer Protocol (FTP) packets with port 20, Kerberos authentication packets with port 88, and so on. Thus in one embodiment, the protocol demux 38 separates all the different protocols in one flow.
In one embodiment, a protocol classifier 40 also sorts the flows in addition to the protocol demux 38. In one embodiment, the protocol classifier 40—operating either in parallel or in sequence with the protocol demux 38—applies signature filters to the flows to attempt to identify the protocol based solely on the transported data. Furthermore, the protocol demux 38 can make a classification decision based on port number which is subsequently overridden by protocol classifier 40. For example, if an individual or program attempted to masquerade an illicit communication (such as file sharing) using an apparently benign port such as port 80 (commonly used for HTTP Web browsing), the protocol classifier 40 would use protocol signatures, i.e., the characteristic data sequences of defined protocols, to verify the speculative classification performed by protocol demux 38.
In one embodiment, the object assembly module 28 outputs each flow organized by protocol, which represent the underlying objects. Referring again to
The object classification module 30 uses the inherent properties and signatures of various documents to determine the content type of each object. For example, a Word document has a signature that is distinct from a PowerPoint document, or an Email document. The object classification module 30 can extract out each individual object and sort them out by such content types. Such classification renders the present invention immune from cases where a malicious user has altered a file extension or other property in an attempt to avoid detection of illicit activity.
In one embodiment, the object classification module 30 determines whether each object should be stored or discarded. In one embodiment, this determination is based on a various capture rules. For example, a capture rule can indicate that Web Traffic should be discarded. Another capture rule can indicate that all PowerPoint documents should be stored, except for ones originating from the CEO's IP address. Such capture rules can be implemented as regular expressions, or by other similar means. Several embodiments of the object classification module 30 are described in more detail further below.
In one embodiment, the capture rules are authored by users of the capture system 22. The capture system 22 is made accessible to any network-connected machine through the network interface module 24 and user interface 34. In one embodiment, the user interface 34 is a graphical user interface providing the user with friendly access to the various features of the capture system 22. For example, the user interface 34 can provide a capture rule authoring tool that allows users to write and implement any capture rule desired, which are then applied by the object classification module 30 when determining whether each object should be stored. The user interface 34 can also provide pre-configured capture rules that the user can select from along with an explanation of the operation of such standard included capture rules. In one embodiment, the default capture rule implemented by the object classification module 30 captures all objects leaving the network 10.
If the capture of an object is mandated by the capture rules, the object classification module 30 can also determine where in the object store module 32 the captured object should be stored. With reference to
Tag Data Structure
In one embodiment, the content store is a canonical storage location, simply a place to deposit the captured objects. The indexing of the objects stored in the content store 44 is accomplished using a tag database 42. In one embodiment, the tag database 42 is a database data structure in which each record is a “tag” that indexes an object in the content store 44 and contains relevant information about the stored object. An example of a tag record in the tag database 42 that indexes an object stored in the content store 44 is set forth in Table 1:
There are various other possible tag fields, and some embodiments can omit numerous tag fields listed in Table 1. In other embodiments, the tag database 42 need not be implemented as a database, and a tag need not be a record. Any data structure capable of indexing an object by storing relational data over the object can be used as a tag data structure. Furthermore, the word “tag” is merely descriptive, other names such as “index” or “relational data store,” would be equally descriptive, as would any other designation performing similar functionality.
The mapping of tags to objects can, in one embodiment, be obtained by using unique combinations of tag fields to construct an object's name. For example, one such possible combination is an ordered list of the Source IP, Destination IP, Source Port, Destination Port, Instance and Timestamp. Many other such combinations including both shorter and longer names are possible. In another embodiment, the tag can contain a pointer to the storage location where the indexed object is stored.
The tag fields shown in Table 1 can be expressed more generally, to emphasize the underlying information indicated by the tag fields in various embodiments. Some of these possible generic tag fields are set forth in Table 2:
For many of the above tag fields in Tables 1 and 2, the definition adequately describes the relational data contained by each field. For the content field, the types of content that the object can be labelled as are numerous. Some example choices for content types (as determined, in one embodiment, by the object classification module 30) are JPEG, GIF, BMP, TIFF, PNG (for objects containing images in these various formats); Skintone (for objects containing images exposing human skin); PDF, MSWord, Excel, PowerPoint, MSOffice (for objects in these popular application formats); HTML, WebMail, SMTP, FTP (for objects captured in these transmission formats); Telnet, Rlogin, Chat (for communication conducted using these methods); GZIP, ZIP, TAR (for archives or collections of other objects); Basic_Source, C++_Source, C_Source, Java_Source, FORTRAN_Source, Verilog_Source, VHDL_Source, Assembly_Source, Pascal_Source, Cobol_Source, Ada_Source, Lisp_Source, Perl_Source, XQuery_Source, Hypertext Markup Language, Cascaded Style Sheets, JavaScript, DXF, Spice, Gerber, Mathematica, Matlab, AllegroPCB, ViewLogic, TangoPCAD, BSDL, C_Shell, K_Shell, Bash_Shell, Bourne_Shell, FTP, Telnet, MSExchange, POP3, RFC822, CVS, CMS, SQL, RTSP, MIME, PDF, PS (for source, markup, query, descriptive, and design code authored in these high-level programming languages); C Shell, K Shell, Bash Shell (for shell program scripts); Plaintext (for otherwise unclassified textual objects); Crypto (for objects that have been encrypted or that contain cryptographic elements); Englishtext, Frenchtext, Germantext, Spanishtext, Japanesetext, Chinesetext, Koreantext, Russiantext (any human language text); Binary Unknown, ASCII Unknown, and Unknown (as catchall categories).
The signature contained in the Signature and Tag Signature fields can be any digest or hash over the object, or some portion thereof. In one embodiment, a well-known hash, such as MD5 or SHA1 can be used. In one embodiment, the signature is a digital cryptographic signature. In one embodiment, a digital cryptographic signature is a hash signature that is signed with the private key of the capture system 22. Only the capture system 22 knows its own private key, thus, the integrity of the stored object can be verified by comparing a hash of the stored object to the signature decrypted with the public key of the capture system 22, the private and public keys being a public key cryptosystem key pair. Thus, if a stored object is modified from when it was originally captured, the modification will cause the comparison to fail.
Similarly, the signature over the tag stored in the Tag Signature field can also be a digital cryptographic signature. In such an embodiment, the integrity of the tag can also be verified. In one embodiment, verification of the object using the signature, and the tag using the tag signature is performed whenever an object is presented, e.g., displayed to a user. In one embodiment, if the object or the tag is found to have been compromised, an alarm is generated to alert the user that the object displayed may not be identical to the object originally captured.
Object Classification
One embodiment of the object classification module 30 is now described in more detail with reference to
For some protocols, such as HTTP, an off-the-shelf protocol handler can be used. For other protocols, the creator of the protocol may provide a protocol handler. Some protocol handlers 62 are designed specially for the capture system 22. In one embodiment, a protocol handlers 62 is included to extract objects from any known transmission protocol, such as HTTP and SMTP. The protocol handlers 62 and object extraction can also be implemented in the object assembly module 28, or in any other module prior to object classification.
Where the object assembly module 28 has been unable to identify the protocol of the flow, the flow is provided to an “unknown protocol handler,” included in the list of protocol handlers 62. In one embodiment, the unknown protocol handler extracts the objects contained in the unidentified flow in the absence of a known protocol. In one embodiment, the unknown protocol handler classifies the entire received flow as a single object. For example, classifying the entire unknown flow as one object can address the difficulty associated with classifying FTP data flows. Other embodiments for the operation of the unknown protocol handler are described further below.
In one embodiment, the extracted object (or objects) is input for the binary signature module 64. As explained above, the binary signature module 64 attempts to classify an object based on binary signatures found inside the object. Binary signatures result from the content encapsulating software operating in some unique manner.
Binary signatures may be inserted on purpose of by happenstance. For example, the binary signature of a Bit Torrent object is the string “BitTorrent” seen at the very beginning of the object. Similarly, all Microsoft Office documents begin with a 32-bit Microsoft identifier based on which each office document can be classified. As another example, JPEG images contain the string “JFIF” at the ninth byte of the object, and the twelfth byte of the object is 0x30 in hexadecimal notation.
Binary signatures may be collected from various sources, such as UNIX “Magic Files,” or additional research and observation. In one embodiment, the signature database containing the signatures of known content types is updated regularly. Signatures can change or become obsolete, while new signatures may be added to known content types or because of new content types.
In one embodiment, if the binary signature module 64 is able to classify the object by content, then the content classification is inserted into the “Content” field of the tag data structure set forth above. If, however, the binary signature module 64 is unable to classify the object, i.e., the object did not match any known signatures, then the object is provided to the object statistics module 66.
In one embodiment, the object statistics module 66 performs various statistical calculations on the object and reaches one or more conclusions based on the results of these calculations. Specifically, in one embodiment, the object statistics module 66 determines whether the object is binary or textual in nature, if binary, whether it is encrypted, and, if textual, what type of text the object contains.
In one embodiment, one statistical analysis performed by the object statistics module 66 calculates the frequency of the bytes contained in the object. In one embodiment, if all 256 possible bytes occur with statistically even frequency, then the object is processed further as a binary object. If, however, certain bytes associated with textual formats—such as ASCII, extended ASCII, or Unicode)—are seen with elevated frequencies, then the object is processed as a text object.
In one embodiment, if the object is determined to be binary data, then the object statistics module 66 performs a distribution analysis (e.g., calculating the variance of the byte distibution) to determine whether the bytes are uniformly distributed or not. In one embodiment, if the bytes are distributed uniformly (to a statistically significant degree), then the object statistics module 66 classifies the object as content type “crypto,” i.e., encrypted data, since most encrypted data appears randomized. In one embodiment, if the byte distribution is found to be non-uniform, the object is classified using the catchall “Binary_Unknown” type. The appropriate classification is then inserted into the tag.
In one embodiment, if the object is determined to be text (e.g., ASCII), then the object statistics module 66 accesses a token database 68 to statistically analyze whether and/or how many times each token appears in the object. A token may be a word, a phrase, a part of a word, grammatical notations, patterns, syntax, and any other textual data. Tokens may vary in size, but will generally be relatively small, usually between 3 and 12 bytes in length. The tokens need not be stored in a token database 68, any appropriate storage scheme and data structure can be used.
The statistical information associated with the tokens is provided, in one embodiment, to the token analyzer 70, which classifies the object as one of a number of various text types using the information. Since various textual documents include different types of syntax, grammar, and words, it is possible to classify text objects using such tokens. For example, certain phrases—such as “is a”, “the”, “and”—appear more regularly in English language text than text in other languages. Similarly, certain tokens—such as “++”, “for (“—appear often in certain programming languages.
In one embodiment, the possible textual content types include Englishtext, Frenchtext, Germantext, Spanishtext, Japanesetext, Chinesetext, Koreantext, Russiantext, (i.e., text from any specific language or a catchall Languagetext category) and various programming language, markup language, query language, and other computer language source code, including Basic_Source, C++_Source, C_Source, Java_Source, FORTRAN_Source, Verilog_Source, VHDL_Source, Assembly_Source, Pascal_Source, Cobol_Source, Ada_Source, Lisp_Source, Perl_Source, XQuery_Source, Hypertext Markup Language, Cascaded Style Sheets, JavaScript, DXF, Spice, Gerber, Mathematica, Matlab, AllegroPCB, ViewLogic, TangoPCAD, BSDL, C_Shell, K_Shell, Bash_Shell, Bourne_Shell, FTP, Telnet, MSExchange, POP3, RFC822, CVS, CMS, SQL, RTSP, MIME, PDF, PS, and Stockdata.
In one embodiment, the tokens in the token database 68 are organized by content type. In other words, each possible content type has tokens associated with it. Furthermore, each token has a numerical weight associated with it. In one embodiment, the token analyzer 70 accesses the token database 68, and calculates a raw number associated with each content type by summing the weights of the tokens found in the object, counting each instance of tokens found more than once. The token analyzer 70 can then classify the object according to the content type with the highest numerical value.
In one embodiment, the tokens in the token database 68 are weighted differently as a function of their frequency and the strength of their association with the specific content type. For example, a common English language word will have a lower weight than a syntax that is highly specific to a certain programming language, such as C++, or other documentation language, such as Verilog.
In one embodiment, the token analyzer 70 assigns a confidence to its classification. For example, if the token summation for Verilog tokens present in an object is twice the total of other content type tokens, then the confidence in a Veriolog classification is relatively high. If, however, two or more content types have token sums that are closer together, the confidence that the content type with the highest token sum is the correct classification is lower. The confidence can be expressed numerically, by ranger, or as a percentage.
In one embodiment, the token analyzer 70 performs object classification using Bayesian statistics, which naturally indicate the probability of the correctness of the classification. For Bayesian analysis the token analyzer only needs to know which tokens are present in the object, but not how many times each token was observed in the object. Based on this input, Bayesian statistics can provide the probability of the object being each of the content types. The highest probability content type can be the classification received by the object.
The probability that the object is the classified content type provided by Bayesian statistics can be converted to, or used as, a confidence in the object classification. In one embodiment, where two (or more) content types are close in probability, both as stored in the content field of the tag, with the appropriate probabilities.
The various embodiments of the object classification method described above have been described in terms of functional modules carrying out the various actions required by each embodiment. However, the modular architecture shown in
The input for object classification remains the captured, assembled, and classified flow of packets. In block 102, a determination is made as to whether the protocol carrying the flow is known. If yes, then in block 104 the appropriate protocol handler associated with the known protocol (e.g., HTTP) is called to extract one or more objects from the flow.
If, on the other hand, the protocol is not determinable or unknown (e.g., an FTP data flow), then in block 106 the unknown protocol handler is called to extract the objects from the flow. In one embodiment, the unknown protocol handler outputs the entire flow as an object. In another embodiment the unknown protocol handler employs methods similar to those discussed with reference to object classification to extract objects from the unclassifiable flow.
In one embodiment, the unknown protocol handler traverses the unknown flow looking for statistically strong binary signatures. If a probable binary signature is found the object embedded in the unknown flow can be simultaneously extracted and classified based on the binary signature without a priori knowledge of the underlying protocol of the flow.
In one embodiment, the unknown protocol handler is configured to identify textual domains—also referred to as ASCII domains—which are regions of the flow identified by a strong ASCII statistical components. If a textual domain is identified in the unclassified flow, the token classification method described above may be employed to extract and classify the textual object content contained in the flow.
After one or more objects are extracted, processing can proceed object by object, or in parallel on a per object basis. In block 108, an attempt is made to classify the object using binary signatures, as set forth above. If in block 110 it is determined that a binary signature has been found, then the object is classified based on the binary signature in block 112 and the processing terminates.
On the other hand, if binary signature classification fails, then in block 114 statistical analysis is performed on the object. This can include, but is not limited to, byte analysis (e.g., how many times each possible byte occurred in the object), byte distribution analysis (e.g., how were the bytes distributed across the object), token presence analysis (e.g., what known tokens were found in the object), and token frequency analysis (e.g., how many times each token was found in the object).
In block 116 a decision is made as to whether the object is binary or textual in nature. For example, if ASCII character bytes occur more frequently than other bytes, the object may be determined to be textual in nature. However, if all bytes occur with approximately even frequency, then the object is probably binary. If the object is binary, then in block 118 a determination is made as to whether the byte distribution is uniform, based on the analysis performed in block 114.
If the distribution of the bytes throughout the object is uniform—defined for example as the variance or standard deviation of the bytes being below three sigma (3 σ) or some other threshold—then the object is classified as a cryptographic object in block 120. In other words, the object is determined to include content that is encrypted by some cryptographic method, and the processing terminates. If, the byte distribution is found to be non-uniform (i.e., non-random), then the object is classified as a binary unknown object in block 122, and the processing terminates.
If, in block 116 the object was determined to be textual in nature, then token analysis is performed on the object in block 124. Token analysis can include calculating totals of token weights found in the object, performing Bayesian statistics of content types based on tokens present in the object, or any other token-based method of determining content type. Based on the calculations performed in block 124, the object is classified as some textual content type in block 126.
In block 128, the confidence of the classification of block 126 is calculated. The confidence may be based on a Bayesian statistic, an comparison of weight sums of other content types, or some other statistical method. The object classification processing then terminates. The object classification derived as a result of the processing can then be used to populate a tag describing the object, or can be associated with the object in some other way, e.g., in a database.
General Matters
In several embodiments, the capture system 22 has been described above as a stand-alone device. However, the capture system of the present invention can be implemented on any appliance capable of capturing and analyzing data from a network. For example, the capture system 22 described above could be implemented on one or more of the servers 14 or clients 16 shown in
In one embodiment, the capture system 22 is an appliance constructed using commonly available computing equipment and storage systems capable of supporting the software requirements. In one embodiment, illustrated by
Thus, a capture system and an object classification procedure have been described. In the forgoing description, various specific values were given names, such as “objects,” and various specific modules, such as the “object statistics module” and “token database” have been described. However, these names are merely to describe and illustrate various aspects of the present invention, and in no way limit the scope of the present invention. Furthermore, various modules, such as the binary signature module 64 and the token analyzer 70 in
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10876205 | Jun 2004 | US |
| Child | 13099516 | US |
