Detecting Backdoors

Yin Zhang and Vern Paxson

Abstract
Backdoors are often installed by attackers who have compromised a system to ease their subsequent return to the system. We consider the problem of identifying a large class of backdoors, namely those providing interactive access on nonstandard ports, by passively monitoring a sites Internet access link. We develop a general algorithm for detecting interactive trafc based on packet size and timing characteristics, and a set of protocol-specic algorithms that look for signatures distinctive to particular protocols. We evaluate the algorithms on large Internet access traces and nd that they perform quite well. In addition, some of the algorithms are amenable to preltering using a stateless packet lter, which yields a major performance increase at little or no loss of accuracy. However, the success of the algorithms is tempered by the discovery that large sites have many users who routinely access what are in fact benign backdoors, such as servers running on non-standard ports not to hide, but for mundane administrative reasons. Hence, backdoor detection also requires a signicant policy component for separating allowable backdoor access from surreptitious access.

dard service such as Telnet, but on an undistinguished port rather than the well-known port associated with the service, or perhaps on a well-known port associated with a different service. In this paper we examine the problem of detecting backdoors, particularly interactive ones, by inspecting network trafc using an intrusion detection system (IDS), where we presume that there is a large volume of legitimate trafc which must be distinguished from the illegitimate trafc. To our knowledge, this problem has not been previously addressed in the literature. Our general approach is to develop a set of algorithms for detecting different types of interactive trafc. These algorithms can then be applied to a trafc stream and whenever they detect interactive trafc using a non-standard service port, we have found some form of backdoor. The rest of the paper is organized as follows. In 2, we discuss the design considerations and examine the tradeoffs of different approaches. In 3, we develop a general algorithm for detecting interactive trafc based on its timing characteristics, and in 4 we present a number of protocol-specic algorithms. In 5, we evaluate the algorithms using traces of Internet trafc. We summarize in 6.

1 Introduction
A backdoor is a mechanism surreptitiously introduced into a computer system to facilitate unauthorized access to the system. While backdoors can be installed for accessing a variety of services, of particular interest for network security are ones that provide interactive access. These are often installed by attackers who have compromised a system to ease their subsequent return to the system. From a network monitoring perspective, such backdoors frequently run over protocols such as Telnet [PR83a], Rlogin [Ka91], or SSH [YKSRL99]. An example of a non-interactive backdoor would be an unauthorized SMTP server [Po82], say to facilitate relaying email spam; and one somewhat in between would be an FTP [PR85] backdoor used to provide access to illicit content such as pirated software, or a Napster server [NA99] run in violation of a sites policy. Backdoors are, by design, difcult to detect. A common scheme for masking their presence is to run a server for a stanY. Zhang is with the Computer Science Department, Cornell University, Ithaca, NY. Email: yzhang@cs.cornell.edu. V. Paxson is with the AT&T Center for Internet Research at ICSI, at the International Computer Science Institute in Berkeley, CA, and with the Lawrence Berkeley National Laboratory. Email: vern@aciri.org. This paper appears in the Proceedings of the 9th USENIX Security Symposium, Denver, Colorado, August 2000.

2 Design Space
A basic principle for backdoor detection is to nd distinctive features indicative of the activity of interest, be it general interactive access, or use of a specic protocol such as SSH. The more powerful a feature is for distinguishing between genuine instances of the activity and false alarms, the better. Candidates for such features include the specic contents of the data stream, the size and transmission rate of the packets in the stream, and their timing structure. This last is potentially very powerful for detecting interactive trafc: studies of Internet trafc have found that the interarrivals of user keystrokes have a striking distribution [DJCME92, PF95], namely a Pareto with innite variance. There is also the possibility that a combination of features will prove to have greater distinctive power than any one feature by itself. We now turn to a discussion of various tradeoffs that arise when considering how to develop detection algorithms.

2.1 Open vs. evasive attackers

In general, network intrusion detection becomes much more difcult when the attacker actively attempts to evade detection by the monitor [PN98, Pa98]. Much of the difculty comes 1

In contrast, timing-based algorithms can be completely unperturbed by the use of encryption. However, timing information can become distorted due to clock skew, propagation delays, loss, and packetization variations. Making timing-based algorithm robust against such noise is challenging.

2.4 Filtering
An important factor for the success of real-time backdoor detection is ltering. The more trafc that can be discarded on a per-packet basis due to patterns in the TCP/IP headers, the better, as this can greatly reduce the processing load on the monitor. As we will see in subsequent sections, ltering can sometimes be highly effective in winnowing down a large trafc stream to just a few packets of interest. However, there is clearly a tradeoff between reduced system load and lost information. First, if a monitor detects suspicious activity in a ltered stream, often the ltering has removed sufcient accompanying context that it becomes quite difcult to determine if the activity is indeed an attack. In addition, the existence of ltering criteria makes it easier for the attackers to evade detection by manipulating their trafc so that it no longer matches the ltering criteria. For example, an evasion against ltering based on packet size (see below) is to use a Telnet client modied to send a large number of do-nothing Telnet options along with each keystroke or line of input. In addition, reliance on ltering can signicantly magnify the problem of chaff, i.e., attackers generating bogus trafc that matches the ltering criteria in order to overwhelm the monitors analysis load, and/or to generate a huge number of false positives, in order to mask a true attack. Three possible ltering criteria for backdoor detection are:

2.2 Passive vs. active monitoring

One tradeoff is whether we only allow the monitor to perform passive monitoring, or if it can actively inject trafc into the network. Passive monitoring has the advantage that it cannot disturb the normal operation of the network. On the other hand, an active monitor could augment its backdoor detection by trying to connect to suspected backdoors in order to probe the server listening on the port to determine its service. However, doing so could in principle tip off the attacker as to the presence of the monitor and the discovery of the backdoor. In this paper we conne ourselves to monitors that only use passive monitoring.

2.3 Content vs. timing

A natural approach for detecting connections to command shell servers is to monitor the keystrokes looking for common shell commands. Such a content-based approach has several drawbacks, however:

[Ra00] for a discussion of experiences with running NFS over email by tunneling IP packets over messages delivered by SMTP.

1 See

Scanning each byte in each incoming packet is very expensive, especially if we must rst reassemble TCP streams to defeat the sort of evasions characterized in [Pa98]. The intruder can then overload the monitor by generating a large amount of legitimate trafc.

from the ability of attackers to exploit ambiguities in a trafc stream. From a monitoring perspective, heuristics might work well for open (non-evasive) attackers, but completely fail in the face of an actively evasive attacker. While ideally any detection algorithms we develop would of course be resistant to evasive attackers, ensuring such robustness can sometimes be exceedingly difcult, and we proceed here on the assumption that there is utility in raising the bar even when a detection algorithm can be defeated by a sufciently aggressive attacker. We further note that if an attacker fully controls both the remote and the local host, and in particular if they are patient and/or able to deploy arbitrary software, then all sorts of devious covert channels become possible1 [Gl93], and backdoor detection becomes essentially hopeless. We do not attempt to address the problem of detecting covert channels. Thus, we propose the algorithms in this paper not as solutions, but merely as waystations in the ongoing arms race between attackers and intrusion detection. One form of arms race we anticipate is particularly likely is between the developers of Napster [NA99] (and Gnutella [GN00]) and our corresponding detection algorithm. Napster has a history of sites attempting to control its use, and of users attempting to circumvent these restrictions [We00], and our algorithm gives sites a new tool for detecting surreptitious use of Napster.

Many command shells allow the user to dene aliases and editing characters, which can easily defeat this approach unless the monitor performs alias and editing expansion of the commands (such as also required for bottleneck analysis [LWWWG98]). Note that this problem can arise either inadvertently, because the attacker as a matter of course uses aliases or redenes the editing sequences, or deliberately, when the attacker is attempting to evade detection. The former case may be amenable to heuristic analysis; the latter likely is not. The intruder can easily evade the monitor by encrypting their content either through some application-level encryption method, or directly using encrypted protocols such as SSH.

Packet size. Keystroke packets are quite small. Even when entire lines of input are transferred using line mode [Bo90], packet payloads will tend to be much smaller than used for bulk-transfer protocols. Therefore, by ltering packets to only capture small packets, the monitor can signicantly reduce its packet capture load.

2.5 Accuracy
As with intrusion detection in general, we face the problem of false positives (non-backdoor connections erroneously agged as backdoors) and false negatives (backdoor connections the monitor fails to detect). The former can make the detection algorithm unusable, because it becomes impossible (or at least too tedious) to examine all of the alerts manually, and attackers can exploit the latter to evade the monitor. We would of course like to have both the false positive rate and the false negative rate be as low as possible. But particularly for those of our algorithms that are based on overall trafc characteristics rather than sharp signatures, we frequently will have to choose tradeoffs between the two.

2.6 Responsiveness
Another important design parameter is the responsiveness of the detection algorithm. That is, after a backdoor connection starts, how long does it take for the monitor to detect the backdoor? Clearly, it is desirable to detect backdoors as quickly as possible, to enable taking additional actions such as recording related trafc or shutting down the connection. However, in many cases waiting longer allows the monitor to gather more information and consequently can detect backdoors more accurately, resulting in a tradeoff of responsiveness versus accuracy. Another consideration related to responsiveness concerns the system resources consumed by the detection algorithm. If we want to detect backdoors quickly, then we must take care not to require more resources than the monitor can devote to detection over a short time period. On the other hand, if off-line analysis is sufcient, then we can use more resourceintensive algorithms. 3

Directionality. In general, an interactive connection such as Telnet is initiated by the client rather than the server, unless the attacker sets up some sort of callback mechanism. This makes it possible to lter connections based on their directionality (inbound vs. outbound). If we are monitoring an Internet access link and are only interested in detecting backdoors at the local site, we can limit our monitoring to just inbound connections, which can signicantly reduce the packet capture load (for example, by ltering out outbound Web surng connections). Note that there is also a cold start problem when the monitor starts running and needs to analyze an existing trafc stream. In this case, it generally cannot determine whether the trafc was initiated inbound or outbound, and accordingly cannot lter it out. Packet contents. When we are interested in identifying specic interactive protocols, it is sometimes possible to lter incoming packets based on patterns specic to the protocol. An example is SSH, discussed in 4.1 below.

3 A General Algorithm for Detecting Interactive Backdoors

In this section we present a general algorithm for detecting interactive backdoors based on keystroke characteristics. The algorithm incorporates three types of characteristics: directionality, packet sizes, and packet interarrival times. We also nd we need to exclude excessively short ows (common in our traces due to the use of scanning by automated monitoring software), which do not provide enough trafc to analyze soundly. The criterion we use is to skip analysis of any ows comprised of fewer than 8 packets or lasting less than 2 seconds, where a ow is one direction of a bidirectional TCP connection.

3.1 Exploiting connection directionality

As noted above, an interactive connection is most likely initiated by the client, unless the server has some callback mechanism. Therefore, when looking for keystrokes we need only consider trafc sent by the initiator of a connection. However, if the monitor doesnt see the establishment of the connection, that is, the connection is a partial connection, there is no way to tell who is the actual initiator. In this case, we must consider both ows. If we are monitoring an access link and are only interested in detecting backdoors within the local site, we can further exploit the connection directionality and ignore all outbound ows, even if the connection is partial.

3.2 Exploiting packet length characteristics

3.2.1 The size of keystroke packets Keystroke packets are likely to be very small, even if sent in line mode, because most commands are short. To verify this assumption, we analyzed several Internet trafc traces with a total of 2.1 million Telnet and Rlogin client data packets. Of these, 79% carried a single byte, 97% carried 3 bytes or less, and 99.7% carried 20 bytes or less. For a trace of SSH 1.x and 2.x connections (very heavily skewed towards 1.x), we found that 28% of the 150 K client data packets had length 20 or less. (Note that those SSH connections with predominantly big packets are likely to be le transfers.) Consequently, we use 20 bytes as our cutoff for small packets. 3.2.2 Characterizing the frequency of small packets Since most keystroke packets are quite small, we can exclude those connections that dont have enough small packets. More specically, we can devise a metric to measure the frequency of small packets in a connection, which we then use to determine whether we should exclude the connection.

3.3 Exploiting timing characteristics

3.4 Making the algorithm run in real-time

In this section we discuss two considerations in using the algorithm in real-time. First, we observe that we can reduce the packet capture load a great deal by ltering on the data 4

As mentioned above, keystroke interarrival times come in a striking Pareto distribution, exhibiting a very broad range [PF95]. We can then exploit the tendency of machinedriven, non-interactive trafc to send packets back-to-back, with a very short interval between them, to discriminate noninteractive trafc from interactive. We do so by examining each pair of back-to-back small packet arrivals and computing the ratio of how many of these interarrival times fall within the range 10 msec through 2 sec. (We need to take care not to include retransmitted packets in this computation.) The upper bound of 2 sec is fairly arbitrary; using 100 sec does not appreciably change the performance. We then dene a metric to quantify how often the interarrival between two consecutive small packets falls in this range. In our nal algorithm, we set the threshold to . It might appear that the criteria of and are too lax, and singularly, they are; but jointly, they prove highly effective, as we show in 5.7.

""4 65&

2 1

The intuition behind is that consecutive small packets are strong indicators that a connection has interactive trafc. If the small packets are all spread throughout a connection, then we will have , so . If they are all grouped together, then and will reect the relative proportion of small packets in the trace. In our nal algorithm, we set the threshold to .

The simplest metric is the ratio of the number of small packets over the total number of packets, for a suitable definition of small packet, which per the previous section we dene as 20 bytes or less of payload. Unfortunately, this metric doesnt work well in practice. Although, as stated in the previous section, over 99.7% of keystrokes are very small, such statistics are based on a large number of connections. For a specic connection, we nd that the ratio can be as low as 3040%. Consequently, in order to prevent frequent false negatives, we have to choose a conservative threshold as low as 2030%. But with such a low threshold, the metrics have little discriminating power and can introduce too many false positives. To avoid such problems, we devised a metric , dened in terms of , the number of small packets, , the total number of packets, and , the number of gaps between small packets. A gap occurs any time two small packets are separated by at least one large packet. We then evaluate:

payload length of the packets to only capture small packets. tcpdump [JLM91] doesnt actually have an easy way to specify a particular range of payload sizes, but the following will lter out all packets with more than 20 bytes of payload: # (packet length # ip header length # tcp header length) <= 20. # That is, data length <= 20. (ip[2:2] - ((ip[0]&0x0f)<<2) (tcp[12]>>2)) <= 20 where the bit-shifting is required to extract the IP and TCP header lengths, which can be variable length due to the presence of IP or TCP options. Introducing ltering does not affect the evaluation of for a ow, since is only computed for packets that are consecutive in the TCP sequence space ( 3.3). However, we must take care when evaluating , since now that we only see small packets, we cant accurately tell the total number of packets transmitted by a given ow. To solve this problem, whenever we see a gap in the sequence number, we estimate the number of missing large packets in the gap as gap LARGE PKT SIZE , where LARGE PKT SIZE is a guess at the most common size for full-sized packets. This size varies with path characteristics such as the Maximum Transmission Unit, and also depends on the particular TCP implementation, but as a rough approximation we simply use . LARGE PKT SIZE The other consideration for real-time detection concerns how quickly the algorithm can determine it has found a backdoor. For off-line analysis, it sufces to check whether a connection has backdoor characteristics when the connection terminates (or when the trace ends), and as we have dened and above, they are in terms of statistics computed over a connections total lifetime. The simplest way to adapt the algorithm to run in real time is to reevaluate and on each incoming packet. Alternatively, we can have a timer for each connection and test the connection whenever the timer goes off. Unfortunately, neither approach works well in practice. The major problem is that when we classify a connection as a non-backdoor connection, we cant just ignore the connection later on, because its hard to tell whether the connection is indeed a non-backdoor connection, or instead actually a backdoor connection with a preamble that has non-backdoor characteristics (such as the Telnet option negotiations that precede a Telnet login dialog). Consequently, we have to keep re-testing each non-backdoor connection, which is clearly very expensive. We address this problem by exponentially backing off the reevaluation timer. We initially choose a small timeout value for the timer (30 seconds). Subsequently, whenever a connection appears to be a non-backdoor, we increase the timeout value by a factor of 1.5, which spreads the computational load over the lifetime of the connection.

"  " ' 0)( ' !)(

" ' &%

" ' 0%

    " #! (

" !$   

4 Special-Purpose rithms

Detection

Algo-

In this section we explore algorithms that look for signatures reecting the use of particular protocols. If we then nd servers for those protocols running on ports other than their standard ones, such instances may indicate the presence of a backdoor. Compared to the general-purpose detection algorithm, special-purpose algorithms can better benet from protocolspecic information, and hence are likely to be more accurate or more efcient. On the other hand, relying on protocolspecic information can make the algorithm susceptible to evasion, if the attacker can perturb the signature. There are two major applications for special-purpose detection algorithms. First, they can be used as baseline algorithms to evaluate the performance of the general-purpose algorithm described in 3, allowing us to understand how much performance we lose by making the algorithm more general (and hence more difcult to evade). Second, the specialpurpose algorithms themselves can be used either individually or in combination with the general-purpose algorithm to detect backdoors. In the rest of this section, we introduce 15 algorithms for detecting various interactive protocols and the like. Based on different design purposes, we can divide these algorithms into the following two classes: Optimal algorithms are designed to identify backdoors as accurately as possible, without worrying about efciency. Such algorithms are intended for use as baseline algorithms and for off-line analysis.

Backdoor type SSH Rlogin Telnet FTP Root prompt Napster Gnutella

Optimal algorithm ssh-sig, ssh-len rlogin-sig telnet-sig ftp-sig root-sig napster-sig gnutella-sig

Efcient algorithm ssh-sig-lter rlogin-sig-lter telnet-sig-lter ftp-sig-lter root-sig-lter napster-sig-lter gnutella-sig-lter

Table 1: Summary of the special-purpose backdoor detection algorithms.

of the form SSH-protoversion-softwareversion comments, followed by carriage-return and newline (ASCII 13 and 10, respectively) [YKSRL99]. The maximum length of the string is 255 characters, including the carriage-return/newline. Version strings contain only printable characters, not including space or -. Currently, the SSH protocol version is either 1.x or 2.x. Therefore, it sufces for ssh-sig to look for text SSH-1. or SSH-2. at the beginning of the rst data packet sent in each direction of a connection. We can replace ssh-sig with the following tcpdump lter (denoted as ssh-sig-lter) for very efcient detection: # 1st 4 bytes are SSH- and # bytes 5 and 6 are 1. or 2. tcp[(tcp[12]>>2):4] = 0x5353482D and (tcp[((tcp[12]>>2)+4):2] = 0x312E or tcp[((tcp[12]>>2)+4):2] = 0x322E) Our second detection algorithm, ssh-len, uses an implicit signature, the packet length, to detect SSH sessions. According to the SSH specication, SSH 1.x will (in the absence of TCP repacketization) generate packet payload sizes of the , that is, 4 more than a multiple of 8. SSH 2.x will form generate payload sizes of length at least 16, and also a multiple of the cipher block size, which is a multiple of 8 for all of the ciphers of which we are aware. Therefore, for SSH, either most packets will have length , or most will have length . One deviation occurs with the initial version exchange, which does not conform with these rules. In light of this pattern, ssh-len detects SSH as follows: 1. First test for an interactive connection using the timingbased algorithm ( 3). If it is interactive, go to the next step, otherwise stop. 2. If the proportion of packets with length or the number of packets with length exceeds a threshold, classify the connection as SSH. We need to be careful when choosing the threshold, because packet retransmission and fragmentation can sometimes distort such characteristics. In our current implementation, we set the threshold to 75%. 5

Table 1 summarizes the algorithms discussed in the rest of this section.

A EC7 9 8

8 C7

Secure Shell (SSH) encrypts transmitted content with strong cryptography. It is increasingly used for both interactive and bulk transfer trafc. While all in all its deployment represents a major advance for Internet security, it presents signicant difculties for content-based intrusion detection precisely because it renders the monitor blind to the specics of each connection. It is thus particularly attractive for backdoor use. Our rst algorithm for detecting SSH, ssh-sig, uses the SSH version string as the signature for SSH. When an SSH connection has been established, both sides send an identifying string

4.1 SSH

A 98 EDC7

A 98 B@67

Efcient algorithms incorporate protocol-specic ltering mechanisms into the optimal algorithms to reduce their expense, at the cost of a degree of accuracy. The tradeoff here varies a great dealsometimes it is even possible to use a simple packet lter to achieve accuracy in the same league as for much more expensive algorithms (see 4.1 below)and the gain is algorithms efcient enough to use for real-time detection.

8 67

4.2 Rlogin
Upon connection establishment, an Rlogin client sends four NUL-terminated strings to the server in the following format [Ka91]: <NUL> client-user-name<NUL> server-user-name<NUL> terminal-type/speed<NUL> The server then returns a zero byte (NUL) to indicate that it has received these strings and is now in data transfer mode. Algorithm rlogin-sig attempts to detect Rlogin sessions using this negotiation as a signature. It rst applies the following analysis to a connection:

IAC IAC IAC IAC

WILL option-code WONT option-code DO option-code DONT option-code

The code values for WILL, WONT, DO, DONT, and IAC are 251, 252, 253, 254, and 255 respectively. Note that some options have parameters, and so can be longer than the above three bytes. telnet-sig tests the rst two bytes of each incoming packet to see if they match the beginning of any of the above. If a connection doesnt involve any option negotiation, we classify it as a non-Telnet connection. Otherwise, we test the following additional conditions:

For the ow sent by the initiator, keep testing each byte until one of the following events happens: - A gap in sequence number occurs; - four NULs have been seen; - an empty string or a non-7-bit-ASCII byte is seen; or - the number of bytes we examined reaches a maximum bound (128 in the current algorithm).

These aid in weeding out binary trafc that happens to match the option patterns above. We can combine the following packet lter with telnet-sig to form a more efcient algorithm, telnet-sig-lter: # 1st byte is <IAC> (0xff), # 2nd byte is <251> - <254> (tcp[(tcp[12]>>2):2] > 0xfffa) and (tcp[(tcp[12]>>2):2] < 0xffff)

If the above terminates by nding four NULs, then we check to see whether the ow in the other direction begins with a non-NUL byte, or whether we found any empty strings or non-7-bit-ASCII bytes. If neither of these last two hold, then the connection is classied as an Rlogin connection. We can combine rlogin-sig with the following tcpdump lter, resulting in a more efcient algorithm rlogin-sig-lter:
# last byte is 0 and data len != 0 and # data length <= 128 (tcp[(ip[2:2]-((ip[0]&0x0f)<<2))-1] = 0) and ((ip[2:2]-((ip[0]&0x0f)<<2)(tcp[12]>>2)) != 0) and ((ip[2:2]-((ip[0]&0x0f)<<2)(tcp[12]>>2)) <= 128)

4.4 FTP
In this section we look at a somewhat different form of interactive protocol, the user control portion of the FTP le transfer protocol [PR85]. FTP is a request/reply protocol in which requests are sent in single, usually short, lines of ASCII text, and replies have a similar structure, but can be longer and multiline. Some FTP requests are sent in response to user activity, and accordingly have interactive-like timing. Others are generated mechanically by the FTP client, and arrive closely spaced. Replies sent by FTP servers start with a status code (a number), followed by any accompanying text. For a days worth of FTP activity between the Lawrence Berkeley National Laboratory and the rest of the Internet (7,229 connections), the distribution of the code in the rst reply returned by the server is: code 220 (ready for new user) seen 6,685 times; code 421 (service not available) seen 535 times; code 226 (closing data connection) seen 7 times; codes 426 (connection closed) and 200 (command okay) each seen once; no other codes seen. Of these, if we miss a server that returns 421 we havent actually missed anything signicant, since the service is not available. All that really matters is detecting 220, though we can include 421, too, without too much extra effort. For FTP server replies, the fourth byte is either a blank or a hyphen, the latter indicating a multi-line reply. Therefore, the ftp-sig algorithm looks in the rst four bytes for either 220 6

Note that rlogin-sig tests for whether the last byte in the packet is NUL, rather than the rst byte. This is necessary because we nd that clients tend to send their rst NUL in its own packet, and the remainder of the prolog information in a second packet.

4.3 Telnet
The Telnet protocol [PR83a] includes a quite general mechanism for negotiating options [PR83b]. Since most Telnet sessions begin with a series of option negotiations, we can attempt to detect these, which have a distinct pattern, taking one of the following four 3-byte formats:

For the ow towards the initiator of a connection, check if the rst byte is a NUL.

At least 75% of the bytes are 7-bit-ASCII. At least 50% of the lines are not longer than 80 bytes.

or 421, followed by either a blank or a hyphen, as a signature for an FTP connection. We can also compose ftp-sig-lter: # 1st three bytes are 220, # 4th byte is blank or hyphen tcp[(tcp[12]>>2):4] = 0x3232302d or tcp[(tcp[12]>>2):4] = 0x32323020 with a similar lter for 421. One difculty with this approach is that the same sort of status codes are used by the popular SMTP mail transfer protocol [Po82]. Code 220 corresponds to service ready and 421 to service not available, just as it does for FTP. This means that our algorithms for detecting FTP backdoors should work just as well for SMTP backdoors (which can actually be benecial), which in 5.5 we explore further.

4.5 Root Backdoor

From operational experience we have found that one particular type of backdoor installed by attackers is a Unix root shell, and the connection to it may not involve any Telnet option negotiation. For these, often the server starts by sending a packet with a payload of exactly two bytes: #<blank>, which corresponds to one of the forms of a Unix root shell prompt. This gives us a simple algorithm, root-sig, which attempts to detect root backdoors by looking for the two bytes in the rst packet sent by the server side of a connection, and the corresponding root-sig-lter: # look for # in a packet with # exactly 2 bytes of payload tcp[(tcp[12]>>2):2] = 0x2320 and (ip[2:2] - ((ip[0]&0x0f)<<2) (tcp[12]>>2)) == 2 which, given its conceptual simplicity, works surprisingly well (see 5.6 below).

servers. Napster users have taken counter-measures to circumvent such blocking [We00], including conguring Napster servers to use non-standard ports for their communications. Open-source Napster clients are also available [GN99, ON00a], which will aid Napster users in modifying the clients behavior to better circumvent detection. Detecting Napster trafc is thus in many ways similar to detecting other backdoors, even though in this case the trafc does not reect a security access violation, but rather a policy violation (authorization rather than authentication). We focused on the problem of detecting the communication directly between Napster clients (used to transfer the actual MP3s). One thought was to develop a generic MP3 detector, though our preliminary work on this has shown the problem to be somewhat difcult, as the format has a short, binary header that does not suggest a simple, distinct pattern to look for [Bo00]. The Napster client communication, however, has a quite distinctive signature [ON00b]. The communication begins with the string SEND or GET followed immediately by the name of the item (no intervening whitespace). Furthermore, we have found that the SEND or GET directive is sent by the Napster client in its own packet,2 so our current version of napster-sig simply looks for either of these strings sent in their own packet and occurring at the beginning of a connection. napster-sig-lter does the same, but without the beginning-of-a-connection context: # look for "SEND" or "GET" in a # packet by itself (so payload of # 4 or 3 bytes, respectively) ((ip[2:2] - ((ip[0]&0x0f)<<2) (tcp[12]>>2)) = 4 and tcp[(tcp[12]>>2):4] = 0x53454e44) or ((ip[2:2] - ((ip[0]&0x0f)<<2) (tcp[12]>>2)) = 3 and tcp[(tcp[12]>>2):2] = 0x4745 and tcp[(tcp[12]>>2)+2]=0x54)

4.6 Napster
Napster is a distributed system by which users can share copies of music that has been digitized in MP3 format [NA99]. Users run a client that connects to napster.com servers for purposes of publishing the MP3s that the user has made available to the public, and for searching for particular MP3s available elsewhere in the distributed database. The server redirects the client to other clients that have the desired MP3 available, and the client then makes a direct connection to the source of the MP3, bypassing the server at this point. Napster has proven controversial because often the MP3 trading is in violation of copyright laws, and also because MP3s tend to be large les, so the enthusiasm of a sites Napster users can consume considerable resources [NA00, Ha00]. Therefore, sites make efforts to control Napster trafc, for example by removing connectivity to the napster.com 7

4.7 Gnutella
Gnutella is a distribution system similar in spirit to Napster [GN00]. Its distinctive features are that it is fully open source, it can be used to exchange arbitrary les and not just MP3s (although there are now Napster add-ons for doing this, too), and it has no centralized componentGnutella clients simply need to know the name of another Gnutella client and they can participate in the distribution network. Consequently, Gnutella is likely to prove harder for sites to control than Napster. In its current form, however, Gnutella is very easy to detect. Each Gnutella session begins with the connecting client transmitting:
2 Clearly, this is very easy for the Napster client to change, and the corresponding change to make to our detector is looking for the absence of whitespace following the directive, which will address mistaking Napster GETs for those used by HTTP.

and receiving in reply:

# look for "GNUTELLA " as first # 9 characters of payload tcp[(tcp[12]>>2):4] = 0x474e5554 and tcp[(4+(tcp[12]>>2)):4] = 0x454c4c41 and tcp[8+(tcp[12]>>2)] = 0x20

5.2 Performance of SSH algorithms

We ran ssh-sig on trace ssh.trace to evaluate its false negative ratio. Clearly, ssh-sig only works when the beginning of a connection is present. Altogether, there are 546 complete SSH connections in ssh.trace, none of which is missed by ssh-sig. This demonstrates that the false negative ratio of ssh-sig is extremely low, which is to be expected since the presence of the signature is required by the specication. We then ran ssh-sig on lbnl.mix1.trace, lbnl.mix2.trace and lbnl.inter.trace to evaluate its false positive ratio. Among the 16,938 complete non-SSH connections, none is mis-classied as SSH by ssh-sig. Therefore, the false positive ratio of ssh-sig is close to 0. ssh-sig-lter has exactly the same good performance on the traces we have, which is not surprising, as the only apparent opportunity for error is unusual packetization splitting the SSH version text across multiple packets. In addition, the ltering gain is tremendous, because only those packets that contain the SSH version string need to be further processed. For ssh.trace, the algorithm needs only inspect 111 KB of packets rather than the 194 MB present in the entire trace. The major limitation of ssh-sig and ssh-sig-lter is that they only work when the beginning of an SSH connection is present. Since SSH can be used for both interactive trafc and bulk transfer, it is difcult to soundly evaluate the false negative ratio of ssh-len, which is designed to detect interactive SSH backdoors. Consequently, we only evaluate the false positive ratio here. Again, we ran ssh-len on the three traces without ssh connections: lbnl.mix1.trace, lbnl.mix2.trace and lbnl.inter.trace. Among the 16,938 non-SSH connections, only 5 are classied as SSH by ssh-len, yielding a very low false positive rate. Compared with ssh-sig and ssh-sig-lter, ssh-len does not require the presence of the beginning of a connection. However, it is less robust for SSH 1.x over highly lossy links, could be coalesced where two SSH blocks of length due to packet retransmission, resulting in a single packet of bytes. Consequently, we only use ssh-len when the beginning of a connection is missing. 8

5 Performance evaluation
In this section we evaluate the algorithms developed in 3 and 4. The evaluations were done by adding implementations of the algorithms to the Bro intrusion detection system [Pa98]. Our general framework for evaluation is as follows. To assess an algorithms accuracy, we rst run it against known interactive trafc of the particular type it is supposed to detect (Telnet, Rlogin, SSH; or, for the general algorithm, a combination of Telnet and Rlogin, since SSH trafc is sometimes bulk-transfer) and analyze how often it fails to ag a connection in the trace as interactive. This evaluates the false negative rate. We then run the algorithm against packet traces of a sites Internet trafc (these have high-volume protocols such as HTTP, NFS, and X11 removed, because otherwise we could not capture the traces reliably) to see which connections they mark as interactive, and then manually assess whether the connection does indeed appear to be interactive. This evaluates the false positive rate. Note, we do not assess the Napster and Gnutella detectors, as the traces we use here were captured before those applications existed. However, our informal assessment based on correlating trafc to known Napster and Gnutella ports and services is that they work very well.

5.1 Trace description

We used four traces to evaluate the performance of the algorithms:

ssh.trace (194MB, 380K packets, 905 connections), a half-hour snapshot of all the SSH connections seen late at night on the Internet access link (DMZ) of the University of California at Berkeley (UCB).

A 9 8 0I67

` 9 V8 9 S P aYXWUTCR8Q7

lbnl.mix1.trace (54MB, 134K packets, 4.6K connections) and lbnl.mix2.trace (421MB, 863K packets, 14.7K connections). Each trace contains one hour of aggregate trafc collected at the DMZ of the

where NL is the newline character (ASCII 10). Accordingly, gnutella-sig looks for the string GNUTELLA blank at the beginning of a connection. The corresponding gnutella-sig-lter is:

F HG

GNUTELLA OK NL

NL

F HG

F HG

GNUTELLA CONNECT/ version

NL

NL

Lawrence Berkeley National Laboratory (LBNL), the rst in the middle of the night, the second in the middle of the afternoon. The traces have had high volume protocols (HTTP, SSH, NFS, X11, NNTP, FTP data) ltered out. Note that we might well apply such ltering for operational use, too, deciding to trade off missing backdoors on those ports for the reduced packet capture load. lbnl.inter.trace (389MB, 3.5M packets, 5.5K connections), one days worth of Telnet and Rlogin trafc collected at LBNL.

5.3 Performance of Rlogin algorithms

Altogether there are 175 complete Rlogin connections in the traces, none of which is missed by rlogin-sig. We begin with evaluating the false positive ratio of rloginsig. In the four traces, altogether there are 17,306 non-rlogin connections, none of which is mis-classied as an Rlogin connection. This means rlogin-sig also has an extremely low false positive ratio. After adding ltering into rlogin-sig, we found that the false negative ratio remains the same (0/175). Meanwhile, the increase in the false positive ratio is marginal: altogether there are 4 out of 17,306 non-Rlogin connections that are misclassied as Rlogin connections by rlogin-sig-lter. The ltering gain of rlogin-sig-lter is signicant. Among the 1 GB data we have in the four traces, only 16 MB data needs to be processed by rlogin-sig. The major limitation of rlogin-sig and rlogin-sig-lter is similar to ssh-sigthey only work when the beginning of a connection is seen by the monitor.

5.5 Performance of FTP algorithms

As noted in 4.4, our FTP detection algorithm will also detect SMTP, so here we note this limitation and then treat the two protocols together. We have altogether 5,629 FTP/SMTP sessions in which the server sent at least 4 bytes of data. Of these, 29 are missed by ftp-sig. Further inspection shows that these connections are almost all partial connections for which the initial dialog (which is far and away the most likely place for our signature to trigger) is missing. This demonstrates that ftp-sig has a low false negative ratio. Among 20,135 non-FTP/SMTP connections, only one is classied as FTP/SMTP. Further inspection shows that this is actually an FTP server running via WinSockso there is no false positive after all! After adding ltering, ftp-sig-lter enjoys the same accuracy, as well as a terric ltering gain: only 1.2 MB out of over 1 GB data need be processed by ftp-sig-lter. Again, the limitation for ftp-sig and ftp-sig-lter is that, except for rare exceptions, they only work when the beginning of a connection is seen by the monitor.

5.4 Performance of Telnet algorithms

Again, we rst evaluate the false negative ratio of algorithm telnet-sig. Unfortunately, it turns out that many Telnet connections in our traces are very short. For such short connections, telnet-sig fails because the connections do not include option negotiations. On the other hand, if a connection is that short, even if it is indeed a backdoor, it is not likely to cause signicant damage. To make the evaluation meaningful, we only consider those connections satisfying:

5.6 Root shell algorithms

As far as we can tell, our traces do not include any root shells, so we cannot soundly evaluate the performance of root-sig and root-sig-lter. But see the next section for preliminary experiences indicating that they (root-sig-lter, in particular) are quite powerful.

the client sends at least two lines of data;

5.7 Performance of the general detection algorithm

To assess the false negative ratio of the algorithm, we ran it on trace lbnl.inter.trace, which consists only of Telnet and Rlogin connections. Among the 150 complete Rlogin connections, 26 are missed by the algorithm. Further inspection shows that 23 are excessively short (less than 2 seconds in duration, or only one command executed), and the other 3 are user login failures. Among all 1,450 Telnet connections that are not excessively short, 22 are missed by the timingbased algorithm. Therefore, the false negative ratio is at least comparable to telnet-sig. Further inspection shows that the algorithm found all 18 connections missed by the telnet-sig, but 22 connections detected by telnet-sig are missed by the timing-based algorithm. To evaluate the false positive ratio of the algorithm, we ran the algorithm on lbnl.mix1.trace and lbnl.mix2.trace with all the Telnet/Rlogin/FTP/SSH/ SMTP connections ltered out. Among over 12,000 connections, the timing-based algorithm reported 57 backdoors. Further inspection shows that 45 are IMAP [Cr94] and POP [MR96] mail servers used interactively, and therefore are not in fact false positives.3
3 The

the server sends at least one line of data; and

the duration of the connection is at least 1 second.

After eliminating connections not satisfying these requirements, 1,526 Telnet connections remain, 18 of which are missed by telnet-sig. Further inspection shows that 17 out of the 18 involve the same public library catalog server, which performs passwordless logins without any option negotiation. We further nd that of the 12,708 non-Telnet connections in the traces, none is mis-classied as Telnet connections. This demonstrates that telnet-sig has a very low false positive ratio. After adding ltering into telnet-sig to form algorithm telnet-sig-lter, the false positive and false negative ratios are unaffected for the traces we have studied. The ltering gain, however, is signicant: telnet-sig-lter has to process less than 1.5 MB out of over 1 GB of packet data. The major limitation of telnet-sig and telnet-sig-lter is similar to ssh-sig and rlogin-sigthey only work when the connection as seen by the monitor includes option negotiations, which tends to only occur at the beginning of a connection. 9

algorithm has also detected interactive SMTP sessions, nominally a

5.8 Experience with production use

We only recently begun operational deployment of the backdoor detection algorithms for production use on the LBNL DMZ. One of the most surprising (and, in retrospect, obvious) ndings has been the large number of legitimate backdoors. For example, when analyzing 20 minutes of trafc from the UCB DMZ (comprising 4.9 GB of data after ltering out the high volume trafc), the protocol-specic algorithms report 334 backdoors on non-standard ports. Of these, 326 are FTP servers on non-standard ports, 7 are interactive games, and the remaining one is a library card catalog server. In contrast, the timing-based algorithm reports 220 backdoors. From visual inspections of 75 of these, we found: 17 are interactive AOL sessions, 19 are interactive games, 14 are chat sessions, 3 are card catalog servers, 7 are FTP sessions, and we were unable to identify the other 15. Running on the live trafc stream, the SSH detection algorithms have turned up SSH servers running on port 80 (nominally HTTPthe server ran on that port to provide tunneling through rewalls); port 110 (nominally POP); port 32 (used to run an older version of SSH than the one on port 22, due to compatibility problems); ports 4432044327 (a NAT server with SSH access to the collection of hosts behind it via a number of different ports); as well as a host of variants of 22 (222, 922, 2222, ). For production use it is unsafe to lter out the high-volume protocols. Running the signature-based tcpdump lters on full trafc streams does not present any performance problems when using a kernel-based packet lter, as the lters are highly selective. For the other protocol-specic detectors, it appears we can also run them on good-sized full trafc streams, as running all of them against a 10 GB trace only takes about 20 CPU minutes on a 400 MHz Pentium II. We run all of the protocol-specic detectors daily against traces of LBNL trafc other than the high-volume ports. (We will shortly be conguring our monitor to run them in realtime.) We currently run with a set of ve lters to remove legitimate backdoors: the NAT front-end mentioned above; two hosts that run a document upload service that triggers ftp-sig (the protocol is not FTP or SMTP, but has a similar structure); a host that runs a service on TCP port 497 that involves an exchange that looks like Telnet option negotiation (but isnt); and a popular FTP server that sometimes serves les with binary data that looks like embedded Telnet options. The Napster and Gnutella detectors have become important tools in enforcing LBNLs appropriate use policy, and, for example, have detected a remote Napster server running on port 21 (FTP) in an apparent attempt to hide or circumvent a rewall. The root backdoor lter, root-sig-lter, has uncovered root backdoors running on UCB trafc. However, these have not been in the form originally intended (in which the connection begins directly with #<blank>), which we know from experience are a rare, albeit striking, signature. Instead, benon-interactive protocol.

cause the lter version of the algorithm detects #<blank> anywhere in a connection, providing it is sent as a prompt (by itself with no newline), root-sig-lter is quite powerful at detecting both some transitions to root via the Unix su command, and sessions for which the prompt seen after the login prolog is indeed #<blank>. Part of the appeal of root-sig-lter is that it generates very few candidate connections, so even though its false hit rate on general trafc is fairly high, the connections it ags are not burdensome to check, and it is an exceptionally cheap algorithm in terms of computation. We do not yet run the general algorithm operationally. As discussed above, it detects large numbers of interactive services, requiring time-consuming effort contacting the managers for the various machines to determine that in fact the backdoors are legitimate. But the potential of the approach seems clear already.

6 Summary
The problem of nding a backdoor connection in a ood of otherwise legitimate network trafc initially appears daunting. But because interactive trafc has characteristics quite different from most machine-driven trafc (smaller packet sizes, longer idle periods), it is possible to search efciently for such trafc. We have presented a general algorithm for doing so, and also protocol-specic algorithms that look for signatures particular to different protocols, both of which we implemented in the Bro intrusion detection system. One unexpected benet of developing the protocol-specic algorithms was to realize how it is frequently possible to ngerprint a particular application protocol by unique or nearly unique text it includes. This lead to the developement of successful algorithms for Napster and Gnutella, which can be important to detect given that their use sometimes violates a sites policy, and that their users often attempt to evade detection. The algorithms are frequently amenable to preltering in which a stateless packet lter discards nearly all of the trafc stream before it is even considered by the algorithm. Such ltering yields major performance increases in terms of reduced CPU processing, for little or sometimes no decrease in accuracy. A related line of future work that may prove fruitful is to explore the possibility of combining the general algorithm with the protocol-specic algorithms, which is likely to yield better accuracy. While the algorithms work very well, a major stumbling block we failed to anticipate is the large number of legitimate backdoors that users routinely access. These are not backdoors in the surreptitious sense, but only in the more general sense of standard protocols being run on non-standard ports. We have recently begun using the algorithms operationally, which will necessitate both the development of rened security policies addressing the many legitimate backdoors, and honing our algorithms as a mechanistic way to eliminate cer10

 cb

tain classes of benign backdoors. But even given these hurdles, we nd the utility of the detection algorithms clear and compelling, and a natural next step is to now investigate their application to detecting custom backdoor protocols such as LOKI [da97] and Back Orice [CERT98].

[Ka91] B. Kantor, BSD Rlogin, RFC 1282, Network Information Center, SRI International, Menlo Park, CA, Dec. 1991. [LWWWG98] R. Lippmann, D. Wyschogrod, S. Webster, D. Weber, and S. Gorton, Using Bottleneck Verication to Find Novel New Attacks with a Low False Alarm Rate, Proc. Recent Advances in Intrusion Detection, Sept. 1998; http://www.zurich.ibm.com/dac/ Prog RAID98/Talks.html#Lippmann 21 . [MR96] J. Myers and M. Rose, Post Ofce Protocol - Version 3, RFC 1939, Network Information Center, DDN Network Information Center, May 1996. [NA99] Napster, http://www.napster.com, 1999. [NA00] Napster (Press Room), http://www.napster.com/ press.html, 2000. [ON00b] Napster protocol specication, http://opennap. sourceforge.net/napster.txt, June 2000. [ON00a] OpenNap, http://opennap.sourceforge.net, 2000. [PF95] V. Paxson and S. Floyd, Wide-Area Trafc: The Failure of Poisson Modeling, IEEE/ACM Transactions on Networking, 3(3), pp. 226-244, June 1995. [Pa98] V. Paxson, Bro: A System for Detecting Network Intruders in Real-Time, Proc. USENIX Security Symposium, Jan. 1998. [Po82] J. Postel, Simple Mail Transfer Protocol, RFC 821, Network Information Center, SRI International, Menlo Park, CA, Aug. 1982. [PR83a] J. Postel and J. Reynolds, Telnet Protocol Specication, RFC 854, Network Information Center, SRI International, Menlo Park, CA, May 1983. [PR83b] J. Postel and J. Reynolds, Telnet Option Specications, RFC 855, Network Information Center, SRI International, Menlo Park, CA, May 1983. [PR85] J. Postel and J. Reynolds, File Transfer Protocol (FTP), RFC 959, Network Information Center, SRI International, Menlo Park, CA, Oct. 1985. [PN98] T. Ptacek and T. Newsham, Insertion, Evasion, and Denial of Service: Eluding Network Intrusion Detection, Secure Networks, Inc., http://www.aciri.org/vern/PtacekNewsham-Evasion-98.ps, Jan. 1998. [Ra00] M. Ranum. RE: Bypassing rewall, mailing list rewall-wizards@nfr.net, Feb. 1, 2000. [We00] D. Weekly, How to get around a Napster blockade, http://david.weekly.org/code/napster-proxy.php3, 2000. [YKSRL99] T. Ylonen, T. Kivinen, M. Saarinen, T. Rinne, and S. Lehtinen, SSH Transport Layer Protocol, Internet Draft, draft-ietf-secsh-transport-07.txt, May 2000.

7 Acknowledgments
We would like to thank Ken Lindahl and Cliff Frost for their greatly appreciated help with gaining research access to UCBs trafc, and Tara Whalen and the anonymous reviewers for their feedback on the work and its presentation.

References
[Bo90] D. Borman, Telnet Linemode Option, RFC 1184, Network Information Center, SRI International, Menlo Park, CA, Oct. 1990. [Bo00] G. Bouvigne, MPEG Audio Layer I/II/III frame header, http://www.mp3-tech.org/programmer/ frame header.html, 2000. [CERT98] CERT Vulnerability Note VN-98.07, http://www. cert.org/vul notes/VN-98.07.backorice.html, Oct 1998. [Cr94] M. Crispin, Internet Message Access Protocol - Version 4, RFC 1730, Network Information Center, DDN Network Information Center, Dec. 1994. [da97] daemon9 route@infonexus.com, LOKI2 (the implementation), Phrack Magazine, 7(51), Sep.01, 1997. http://www.infowar.com/iwftp/phrack/Phrack51/P5106.txt . [DJCME92] P. Danzig, S. Jamin, R. C ceres, D. Mitzel, and a D. Estrin, An Empirical Workload Model for Driving Wide-area TCP/IP Network Simulations, Internetworking: Research and Experience, 3(1), pp. 1-26, 1992. [Gl93] V. Gligor, A Guide to Understanding Covert Channel Analysis of Trusted Systems, NCSC-TG-030, version 1, http://www.radium.ncsc.mil/tpep/lib-rary/rainbow/ NCSC-TG-030.html, National Computer Security Center, Nov. 1993. [GN99] Gnapster, gnapster.html, 1999. http://www.faradic.net/ jasta/

[GN00] Gnutella, http://gnutella.wego.com, 2000. [Ha00] J. Harrow, The Consumer Internet Steamroller, The Rapidly Changing Face of Computing, http://www. compaq.com/rcfoc/20000417.html# Toc480185377, April, 2000. [JLM91] V. Jacobson, C. Leres, and S. McCanne, tcpdump, ftp://ftp.ee.lbl.gov/tcpdump.tar.Z, 1991. 11