1. Abstract
HTTP Fingerprinting is a relatively new topic of discussion in the context of application security. One of the biggest challenges of maintaining a high level of network security is to have a complete and accurate inventory of networked assets. Web servers and web applications have now become a part of the scope of a network security assessment exercise. In this paper, we present techniques to identify various types of HTTP servers. We shall discuss some of the problems faced in inventorying HTTP servers and how we can overcome them.
We shall also introduce and describe a tool, httprint, which is built using the concepts discussed in this paper.
2. Theory of Fingerprinting
A fingerprint is defined as:
The impression of a fingertip on any surface; also: an ink impression of the lines upon the fingertip taken for the purpose of identification.
something that identifies: as (a) a trait, trace, or characteristic revealing origin or responsibility (b) analytical evidence that characterizes an object or substance.
The process of fingerprinting can be broken up into two sub processes, namely gathering and classification of fingerprints, and comparision of unknown fingerprints with the stored database of known fingerprints.
While gathering fingerprints, it is essential to capture all the key characteristics of the object revealed in the fingerprint. Capturing more details and traits helps in the comparision process. While comparing fingerprints, there may be chances that a fingerprint can be improperly matched, because of subtle differences that can be easily mistaken.
Fingerprinting is a known technique in network security. Operating system fingerprinting is a common task in any network assessment or inventorying exercise. There are many techniques to perform operating system fingerprinting. What makes operating system fingerprinting successful and accurate is the fact that each operating system implements the TCP/IP stack slightly differently. The way a system responds to malformed packets, either the presence of an error response, or absence thereof, is one example of an implementation difference. A detailed discussion on operating system fingerprinting, or TCP/IP stack fingerprinting, is presented in Fyodor's paper, titled "Remote OS detection via TCP/IP Stack Fingerprinting" [1]
The theory behind HTTP fingerprinting is more or less on the same lines - identifying HTTP servers by their implementation differences in the HTTP protocol. HTTP fingerprinting gets slightly more complicated than TCP/IP stack fingerprinting. The reason being that it is easily possible to customize the responses of an HTTP server by just changing its configuration file, or adding plug-ins or modules, whereas customising the behaviour of the TCP/IP stack requires access to the network code at the kernel layer. Despite this difficulty, it is possible to devise tests to overcome the various customizable features of a web server.
3. Banner grabbing
The simplest and most basic form of identifying HTTP servers is to look at the Server field in the HTTP response header [2]. Using a TCP client like netcat [3], it is possible to send an HTTP request to return the HTTP response header of the server, as shown below:
$ nc 202.41.76.251 80HEAD / HTTP/1.0HTTP/1.1 200 OKDate: Mon, 16 Jun 2003 02:53:29 GMTServer: Apache/1.3.3 (Unix) (Red Hat/Linux)Last-Modified: Wed, 07 Oct 1998 11:18:14 GMTETag: "1813-49b-361b4df6"Accept-Ranges: bytesContent-Length: 1179Connection: closeContent-Type: text/html$
Three examples of the HTTP response header are shown below:
From an Apache 1.3.23 server:
HTTP/1.1 200 OKDate: Sun, 15 Jun 2003 17:10:49 GMTServer: Apache/1.3.23Last-Modified: Thu, 27 Feb 2003 03:48:19 GMTETag: "32417-c4-3e5d8a83"Accept-Ranges: bytesContent-Length: 196Connection: closeContent-Type: text/html
From a Microsoft IIS 5.0 server:
HTTP/1.1 200 OKServer: Microsoft-IIS/5.0Expires: Tue, 17 Jun 2003 01:41:33 GMTDate: Mon, 16 Jun 2003 01:41:33 GMTContent-Type: text/htmlAccept-Ranges: bytesLast-Modified: Wed, 28 May 2003 15:32:21 GMTETag: "b0aac0542e25c31:89d"Content-Length: 7369
From a Netscape Enterprise 4.1 server:
HTTP/1.1 200 OKServer: Netscape-Enterprise/4.1Date: Mon, 16 Jun 2003 06:19:04 GMTContent-type: text/htmlLast-modified: Wed, 31 Jul 2002 15:37:56 GMTContent-length: 57Accept-ranges: bytesConnection: close
4. Applications of HTTP fingerprinting
From a network management standpoint, HTTP fingerprinting comes in very handy when keeping track of the various web servers on a network. HTTP fingerprinting can also be used to automate information systems and security audits. Automated security testing tools can use HTTP fingerprinting to narrow down the set of tests required, based on the specific platform or the specific web server being audited.
Some of the applications of HTTP fingerprinting are:
Network management: Web server inventory
Penetration testing / Auditing: Selecting the right attacks or audit tests for web servers
Wireless networks: Detecting 802.11x access points from the wired network, since most APs have a web enabled interface
Web enabled devices: Printers, storage servers, switches, etc. Many times, web enabled devices do not return a server banner string at all, making it difficult to identify and track them.
5. Obfuscating the server banner string
Banner grabbing proves to be a good method of HTTP fingerprinting in many cases. However, many times, server administrators chose to disguise the server banner string, by providing one of their own. Such, "security-by-obscurity" methods help thwart a lot of automated attacks against web servers.
It is easy to configure servers to return different server banner strings. In open source servers such as Apache, one can change the banner string in the source code and re-compile the server. For non-open source servers such as Microsoft IIS or Netscape Enterprise, it is possible to "patch" the binary by opening it up in a hex editor and changing the string embedded in the binary. It is not so easy to do this always, but it has been done successfully. Another way of obscuring the server banner string is to write a custom plug-in for the web server, which can provide customized HTTP responses. A commercial product, called ServerMask [4] from Port 80 Software performs such obfuscation on HTTP responses.
The example below shows the response from an HTTP server with a customized server banner string:
Apache Server recompiled with "Unknown-Webserver/1.0" as the server banner string
HTTP/1.1 403 ForbiddenDate: Mon, 16 Jun 2003 02:41:27 GMTServer: Unknown-Webserver/1.0Connection: closeContent-Type: text/html; charset=iso-8859-1
The example below shows the response from an HTTP server using ServerMask:
IIS Server using the ServerMask plug-in
HTTP/1.1 200 OKServer: Yes we are using ServerMaskDate: Mon, 16 Jun 2003 02:54:17 GMTConnection: Keep-AliveContent-Length: 18273Content-Type: text/htmlSet-Cookie: It works on cookies too=82.3S3.O12.NT2R0RE,4147ON3P,.4OO.; path=/Cache-control: private
As we can see from the above examples, relying purely upon the contents of the server banner string is not enough for identifying the type of HTTP server.
6. Protocol behaviour
Almost all HTTP servers differ in the way they implement the HTTP protocol. In the case where the HTTP request is well formed and legitimate, the response returned by all HTTP servers is more or less compliant with the specifications laid out in the RFCs for HTTP[5]. However, when confronted with malformed HTTP requests, these servers differ in their responses. Differences in the way the HTTP protocol is handled by various HTTP servers forms the basis of the HTTP fingerprinting technique.
Let us illustrate these differences with examples. We shall analyse the response to four HTTP requests, coming from an Apache 1.3.23 server, a Microsoft IIS 5.0 server and a Netscape Enterprise 4.1. The requests are:
HTTP Test What to expect
HEAD / HTTP/1.0 Normal HTTP header response
DELETE / HTTP/1.0 Response when operations such as DELETE aren't generally allowed
GET / HTTP/3.0 Response to a request with an improper HTTP protocol number
GET / JUNK/1.0 Response to a request with an improper protocol specification
In each of these responses, we shall identify key differences between the responses of Apache 1.3.23, IIS 5.0 and Netscape Enterprise 4.1. We shall not take into consideration differences in customizable parameters such as the server banner string.
6.1 HTTP header field ordering
Taking the first request HEAD / HTTP/1.0, we shall analyse the HTTP response header and inspect the order of appearance of the various fields returned within it.
Response from Apache 1.3.23
$ nc apache.example.com 80HEAD / HTTP/1.0HTTP/1.1 200 OKDate: Sun, 15 Jun 2003 17:10:49 GMTServer: Apache/1.3.23 Last-Modified: Thu, 27 Feb 2003 03:48:19 GMTETag: "32417-c4-3e5d8a83"Accept-Ranges: bytesContent-Length: 196Connection: closeContent-Type: text/html
Response from IIS 5.0
$ nc iis.example.com 80HEAD / HTTP/1.0HTTP/1.1 200 OKServer: Microsoft-IIS/5.0Content-Location: http://iis.example.com/Default.htmDate: Fri, 01 Jan 1999 20:13:52 GMTContent-Type: text/htmlAccept-Ranges: bytesLast-Modified: Fri, 01 Jan 1999 20:13:52 GMTETag: W/"e0d362a4c335be1:ae1"Content-Length: 133
Response from Netscape Enterprise 4.1
$ nc netscape.example.com 80HEAD / HTTP/1.0HTTP/1.1 200 OKServer: Netscape-Enterprise/4.1Date: Mon, 16 Jun 2003 06:01:40 GMTContent-type: text/htmlLast-modified: Wed, 31 Jul 2002 15:37:56 GMTContent-length: 57Accept-ranges: bytesConnection: close
If we observe the ordering of the response header fields Server and Date, we notice that Apache orders the fields differently than IIS and Netscape.
6.2 HTTP DELETE (forbidden operation) response
Next, we shall take the request DELETE / HTTP/1.0 and observe what the response of each of the servers is, when the requested operation is generally forbidden.
Response from Apache 1.3.23
$ nc apache.example.com 80DELETE / HTTP/1.0HTTP/1.1 405 Method Not AllowedDate: Sun, 15 Jun 2003 17:11:37 GMTServer: Apache/1.3.23 Allow: GET, HEAD, POST, PUT, DELETE, CONNECT, OPTIONS, PATCH, PROPFIND, PROPPATCH, MKCOL, COPY, MOVE, LOCK, UNLOCK, TRACEConnection: closeContent-Type: text/html; charset=iso-8859-1
Response from IIS 5.0
$ nc iis.example.com 80DELETE / HTTP/1.0HTTP/1.1 403 ForbiddenServer: Microsoft-IIS/5.0Date: Fri, 01 Jan 1999 20:13:57 GMTContent-Type: text/htmlContent-Length: 3184
Response from Netscape Enterprise 4.1
$ nc netscape.example.com 80DELETE / HTTP/1.0HTTP/1.1 401 UnauthorizedServer: Netscape-Enterprise/4.1Date: Mon, 16 Jun 2003 06:03:18 GMTWWW-authenticate: Basic realm="WebServer Server"Content-length: 223Content-type: text/htmlConnection: close
Apache responds with a 405 "Method not allowed" response, IIS responds with a 403 "Operation on resource forbidden" response, and Netscape responds with a 401 "Authorization credentials required" response. Each of the servers differs in their response to the DELETE request.
6.3 Improper HTTP version response
The next test consists of sending an HTTP request with an improper HTTP version number, such as GET / HTTP/3.0, to the server. HTTP 3.0 is not even in existence as of this writing, and none of the candidate servers implement it.
Response from Apache 1.3.23
$ nc apache.example.com 80GET / HTTP/3.0HTTP/1.1 400 Bad RequestDate: Sun, 15 Jun 2003 17:12:37 GMTServer: Apache/1.3.23 Connection: closeTransfer-Encoding: chunkedContent-Type: text/html; charset=iso-8859-1
Response from IIS 5.0
$ nc iis.example.com 80GET / HTTP/3.0HTTP/1.1 200 OKServer: Microsoft-IIS/5.0Content-Location: http://iis.example.com/Default.htmDate: Fri, 01 Jan 1999 20:14:02 GMTContent-Type: text/htmlAccept-Ranges: bytesLast-Modified: Fri, 01 Jan 1999 20:14:02 GMTETag: W/"e0d362a4c335be1:ae1"Content-Length: 133
Response from Netscape Enterprise 4.1
$ nc netscape.example.com 80GET / HTTP/3.0HTTP/1.1 505 HTTP Version Not SupportedServer: Netscape-Enterprise/4.1Date: Mon, 16 Jun 2003 06:04:04 GMTContent-length: 140Content-type: text/htmlConnection: close
Apache responds with a 400 "Bad HTTP request" response, IIS ignores the improper HTTP protocol number, and responds with a 200 "OK" along with the contents of the HTML data for the root document, and Netscape responds with a 505 "HTTP version not supported" response.
6.4 Improper protocol response
The next test involves observing the response to the request GET / JUNK/1.0.
Response from Apache 1.3.23
$ nc apache.example.com 80GET / JUNK/1.0HTTP/1.1 200 OKDate: Sun, 15 Jun 2003 17:17:47 GMTServer: Apache/1.3.23 Last-Modified: Thu, 27 Feb 2003 03:48:19 GMTETag: "32417-c4-3e5d8a83"Accept-Ranges: bytesContent-Length: 196Connection: closeContent-Type: text/html
Response from IIS 5.0
$ nc iis.example.com 80GET / JUNK/1.0HTTP/1.1 400 Bad RequestServer: Microsoft-IIS/5.0Date: Fri, 01 Jan 1999 20:14:34 GMTContent-Type: text/htmlContent-Length: 87
Response from Netscape Enterprise 4.1
$ nc netscape.example.com 80GET / JUNK/1.0
Bad request
Your browser sent a query this server could not understand.In this case, Apache ignores the improper protocol "JUNK", and responds with a 200 "OK" along with the contents of the root document, IIS responds with a 400 "Bad Request" response and Netscape does not even return an HTTP response header, but instead just returns an HTML formatted error message stating that this request is a bad request.
6.5 Summary of test results
The following table summarizes the various tests and the responses from each of the HTTP servers. It is easy to figure out how to distinguish HTTP servers from such tests.
Server Field Ordering DELETE Method Improper HTTP version Improper protocol
Apache/1.3.23 Date, Server 405 400 200
Microsoft-IIS/5.0 Server, Date 403 200 400
Netscape-Enterprise/4.1 Server, Date 401 505 no header
6.6 Choosing the right tests
In the above example, we discussed four HTTP tests, and observed the differences in the responses from three popular HTTP servers. For an industrial strength HTTP fingerprinting engine, we require more than four HTTP tests. The larger the number of HTTP tests, the better the results of fingerprinting, and the greater the accuracy in matching fingerprints. On the other hand, fewer HTTP tests imply faster execution time.
Fingerprinting tests are of two types
Decision tree bases tests
Statistical analysis tests
Decision tree based tests rely on the construction of a tree of tests, which eliminates possibilities progressively, much in the same manner as testing for an unknown chemical compound, based on the results of progressive chemical reaction tests. Decision tree based tests are difficult to scale, and each HTTP server would have specific contributions to the construction of the decision tree. Adding tests for a new HTTP server would involve re-writing the entire decision tree.
Statistical analysis tests usually involve a fixed set of tests, which result in an array of weights being generated for each type of HTTP server. The decision as to what the outcome is, is based on comparing the various weights generated for each server. The accuracy of statistical analysis tests depends on the algorithms used to assign and compare the weights for each HTTP server. Statistical based models yield themselves quite easily to be adapted into neural machines, which can be trained with a set of known values.
7. Statistical and Fuzzy analysis
The rest of the paper focusses on using statistical and fuzzy logic techniques in analysing the responses from the HTTP tests. The technique consists in perfomrming signature analysis with a set of stored signatures, and assigning a confidence rating to each candidate signature. The signatures with the highest confidence rating are then reported as potential matches for the unknown server being tested.
7.1 Assumptions
The fingerprinting engine operates with a known set of server signatures. It can therefore only identify HTTP servers that it knows about. If a server's signature is not present in the set of known signatures, the fingerprinting engine shall report the next closest server, in terms of server behaviour and characteristics.
While performing HTTP fingerprint tests, there shall be no HTTP proxy server present between the system running the fingerprinting engine and the target web server.
7.2 Terms and Definitions
Term Description
Signature Set S = {s1, s2, ..., sn}
n = number of web server signatures known to the fingerprinting engine.
si = ith signature in the signature set.
Reported Signature sR
The signature that is derived from running HTTP fingerprinting tests against an unknown web server.
Comparision function
and Weight wi = fw(sR, si)
wi = Weight when reported signature sR is compared against the ith signature in the signature set S.
fw(sa, sb) = Comparision function, which compares signature sa against sb, and returns a resultant weight.
Weight Vector W = {w1, w2, ..., wn}
Confidence Rating ci = fc(wi, W)
ci = likelihood that signature si, with weight wi is the best match amongst the entire signature set S, whose weight vector is represented by W.
fc(wa, W) = Fuzzy logic function to calculate the likelihood, in percentage terms, of wa being the best weight amongst the weight vector W.
Confidence Vector C = {c1, c2, ..., cn}
Max Confidence cmax
Maximum confidence rating amongst all the confidence ratings in the confidence vector C.
Best match Vector M = {smaxA, smaxB, ...}
smaxA, smaxB = signatures whose confidence ratings equal to cmax
7.3 Analysis Logic
The following piece of pseudo code outlines how the best match is chosen out of the signature set.
Load the signature set S = {s1, ..., sn}
Run the fingerprinting tests and obtain sR for the unknown web server.
for i = 1 .. n
wi = fw(sR, si)
next
for i = 1 .. n
ci = fc(wi, W)
next
cmax = max(C)
M = {}
for i = 1 .. n
if ci = cmax then
M = M U {si}
end if
next
print M
8. httprint - the advanced HTTP fingerprinting engine
The httprint [6] fingerprinting engine uses statistical analysis, combined with fuzzy logic techniques, to determine the type of HTTP server.
httprint can be used to both gather as well as analyse signatures generated from HTTP servers.
Although httprint is not open source, it is available at no cost for personal, educational and non-commercial use.
8.1 httprint signatures
HTTP signatures generated by httprint are hexadecimal encoded ASCII strings, as the ones shown below:
Microsoft-IIS/5.0CD2698FD6ED3C295E4B1653082C10D64050C5D2594DF1BD04276E4BB811C9DC50D7645B5811C9DC52A200B4C9D69031D6014C217811C9DC5811C9DC52655F350FCCC535BE2CE6923E2CE69232FCD861AE2CE69272576B769E2CE6926CD2698FD6ED3C295E2CE692009DB9B3E811C9DC5811C9DC56ED3C2956ED3C295E2CE69236ED3C2956ED3C295811C9DC5E2CE69276ED3C295Apache/2.0.x9E431BC86ED3C295811C9DC5811C9DC5050C5D32505FCFE84276E4BB811C9DC50D7645B5811C9DC5811C9DC5CD37187C11DDC7D7811C9DC5811C9DC58A91CF57FCCC535B6ED3C295FCCC535B811C9DC5E2CE6927050C5D336ED3C2959E431BC86ED3C295E2CE69262A200B4C6ED3C2956ED3C2956ED3C2956ED3C295E2CE6923E2CE69236ED3C295811C9DC5E2CE6927E2CE6923
httprint maintains a set of signatures in a text file [7], and uses these to analyse the results generated from an unknown server. It is easily possible to extend the signatures database, by simply cutting-and-pasting the signature output of httprint, when used against a known server whose fingerprint is not in the database. The next time httprint is run, the newly added signature will be used in the comparision.
8.2 httprint command line and GUI interfaces
httprint is available in both command-line and GUI versions, running on Windows, Linux, Mac OS X and FreeBSD for this release - v200.
httprint's command line options are as under:
# ./httprint httprint v0.200 (beta) - web server fingerprinting tool(c) 2003, net-square solutions pvt. ltd. - see readme.txthttp://net-square.com/httprint/
httprint@net-square.comUsage:httprint {-h
No comments:
Post a Comment