Intrusion Detection Systems
After the perimeter controls, firewall, and authentication and access controls block certain actions, some users are admitted to use a computing system. Most of these controls are preventive: they block known bad things from happening. Many studies (for example, see [DUR99]) have shown that most computer security incidents are caused by insiders, people who would not be blocked by a firewall. And insiders require access with significant privileges to do their daily jobs. The vast majority of harm from insiders is not malicious; it is honest people making honest mistakes. Then, too, there are the potential malicious outsiders who have somehow passed the screens of firewalls and access controls. Prevention, although necessary, is not a complete computer security control; detection during an incident copes with harm that cannot be prevented in advance. Halme and Bauer [HAL95] survey the range of controls to address intrusions.
Intrusion detection systems complement these preventive controls as the next line of defense. An intrusion detection system (IDS) is a device, typically another separate computer, that monitors activity to identify malicious or suspicious events. An IDS is a sensor, like a smoke detector, that raises an alarm if specific things occur. A model of an IDS is shown in Figure 7-42. The components in the figure are the four basic elements of an intrusion detection system, based on the Common Intrusion Detection Framework of [STA96]. An IDS receives raw inputs from sensors. It saves those inputs, analyzes them, and takes some controlling action.
Figure 7-42 Common Components of an Intrusion Detection Framework.
IDSs perform a variety of functions:
monitoring users and system activity
auditing system configuration for vulnerabilities and misconfigurations
assessing the integrity of critical system and data files
recognizing known attack patterns in system activity
identifying abnormal activity through statistical analysis
managing audit trails and highlighting user violation of policy or normal activity
correcting system configuration errors
installing and operating traps to record information about intruders
No one IDS performs all of these functions. Let us look more closely at the kinds of IDSs and their use in providing security.
Types of IDSs
The two general types of intrusion detection systems are signature based and heuristic. Signature-based intrusion detection systems perform simple pattern-matching and report situations that match a pattern corresponding to a known attack type. Heuristic intrusion detection systems, also known as anomaly based, build a model of acceptable behavior and flag exceptions to that model; for the future, the administrator can mark a flagged behavior as acceptable so that the heuristic IDS will now treat that previously unclassified behavior as acceptable.
Intrusion detection devices can be network based or host based. A network-based IDS is a stand-alone device attached to the network to monitor traffic throughout that network; a host-based IDS runs on a single workstation or client or host, to protect that one host.
Early intrusion detection systems (for example, [LUN90b, FOX90]) worked after the fact, by reviewing logs of system activity to spot potential misuses that had occurred. The administrator could review the results of the IDS to find and fix weaknesses in the system. Now, however, intrusion detection systems operate in real time (or near real time), watching activity and raising alarms in time for the administrator to take protective action.
Signature-Based Intrusion Detection
A simple signature for a known attack type might describe a series of TCP SYN packets sent to many different ports in succession and at times close to one another, as would be the case for a port scan. An intrusion detection system would probably find nothing unusual in the first SYN, say, to port 80, and then another (from the same source address) to port 25. But as more and more ports receive SYN packets, especially ports that are not open, this pattern reflects a possible port scan. Similarly, some implementations of the protocol stack fail if they receive an ICMP packet with a data length of 65535 bytes, so such a packet would be a pattern for which to watch.
The problem with signature-based detection is the signatures themselves. An attacker will try to modify a basic attack in such a way that it will not match the known signature of that attack. For example, the attacker may convert lowercase to uppercase letters or convert a symbol such as "blank space" to its character code equivalent %20. The IDS must necessarily work from a canonical form of the data stream in order to recognize that %20 matches a pattern with a blank space. The attacker may insert malformed packets that the IDS will see, to intentionally cause a pattern mismatch; the protocol handler stack will discard the packets because of the malformation. Each of these variations could be detected by an IDS, but more signatures require additional work for the IDS, which reduces performance.
Of course, signature-based IDSs cannot detect a new attack for which a signature is not yet installed in the database. Every attack starts as a new attack at some time, and the IDS is helpless to warn of its existence.
Ideally, signatures should match every instance of an attack, match subtle variations of the attack, but not match traffic that is not part of an attack. However, this goal is grand but unreachable.
Heuristic Intrusion Detection
Because signatures are limited to specific, known attack patterns, another form of intrusion detection becomes useful. Instead of looking for matches, heuristic intrusion detection looks for behavior that is out of the ordinary. The original work in this area (for example, [TEN90]) focused on the individual, trying to find characteristics of that person that might be helpful in understanding normal and abnormal behavior. For example, one user might always start the day by reading e-mail, write many documents using a word processor, and occasionally back up files. These actions would be normal. This user does not seem to use many administrator utilities. If that person tried to access sensitive system management utilities, this new behavior might be a clue that someone else was acting under the user's identity. The approach has been extended to networks in [MUK94]. Later work (for example, [FOR96]) sought to build a dynamic model of behavior, to accommodate variation and evolution in a person's actions over time. The technique compares real activity with a known representation of normality.
Alternatively, intrusion detection can work from a model of known bad activity. For example, except for a few utilities (login, change password, create user), any other attempt to access a password file is suspect. This form of intrusion detection is known as misuse intrusion detection. In this work, the real activity is compared against a known suspicious area.
All heuristic intrusion detection activity is classified in one of three categories: good/benign, suspicious, or unknown. Over time, specific kinds of actions can move from one of these categories to another, corresponding to the IDS's learning whether certain actions are acceptable or not.
As with pattern-matching, heuristic intrusion detection is limited by the amount of information the system has seen (to classify actions into the right category) and how well the current actions fit into one of these categories.
An IDS is a network device (or, in the case of a host-based IDS, a program running on a network device). Any network device is potentially vulnerable to network attacks. How useful would an IDS be if it itself were deluged with a denial-of-service attack? If an attacker succeeded in logging in to a system within the protected network, wouldn't trying to disable the IDS be the next step?
To counter those problems, most IDSs run in stealth mode, whereby an IDS has two network interfaces: one for the network (or network segment) being monitored and the other to generate alerts and perhaps other administrative needs. The IDS uses the monitored interface as input only; it never sends packets out through that interface. Often, the interface is configured so that the device has no published address through the monitored interface; that is, a router cannot route anything to that address directly, because the router does not know such a device exists. It is the perfect passive wiretap. If the IDS needs to generate an alert, it uses only the alarm interface on a completely separate control network. Such an architecture is shown in Figure 7-43.
Figure 7-43 Stealth Mode IDS Connected to Two Networks.
Other IDS Types
Some security engineers consider other devices to be IDSs as well. For instance, to detect unacceptable code modification, programs can compare the active version of a software code with a saved version of a digest of that code. The tripwire program [KIM98] is the most well known software (or static data) comparison program. You run tripwire on a new system, and it generates a hash value for each file; then you save these hash values in a secure place (offline, so that no intruder can modify them while modifying a system file). If you later suspect your system may have been compromised, you rerun tripwire, providing it the saved hash values. It recomputes the hash values and reports any mismatches, which would indicate files that were changed.
System vulnerability scanners, such as ISS Scanner or Nessus, can be run against a network. They check for known vulnerabilities and report flaws found.
As we have seen, a honeypot is a faux environment intended to lure an attacker. It can be considered an IDS, in the sense that the honeypot may record an intruder's actions and even attempt to trace who the attacker is from actions, packet data, or connections.
Goals for Intrusion Detection Systems
The two styles of intrusion detectionpattern-matching and heuristicrepresent different approaches, each of which has advantages and disadvantages. Actual IDS products often blend the two approaches.
Ideally, an IDS should be fast, simple, and accurate, while at the same time being complete. It should detect all attacks with little performance penalty. An IDS could use someor allof the following design approaches:
- filter on packet headers
- filter on packet content
- maintain connection state
- use complex, multipacket signatures
- use minimal number of signatures with maximum effect
- filter in real time, online
- hide its presence
use optimal sliding time window size to match signatures
Responding to Alarms
Whatever the type, an intrusion detection system raises an alarm when it finds a match. The alarm can range from something modest, such as writing a note in an audit log, to something significant, such as paging the system security administrator. Particular implementations allow the user to determine what action the system should take on what events.
What are possible responses? The range is unlimited and can be anything the administrator can imagine (and program). In general, responses fall into three major categories (any or all of which can be used in a single response):
monitor, collect data, perhaps increase amount of data collected
protect, act to reduce exposure
call a human
Monitoring is appropriate for an attack of modest (initial) impact. Perhaps the real goal is to watch the intruder, to see what resources are being accessed or what attempted attacks are tried. Another monitoring possibility is to record all traffic from a given source for future analysis. This approach should be invisible to the attacker. Protecting can mean increasing access controls and even making a resource unavailable (for example, shutting off a network connection or making a file unavailable). The system can even sever the network connection the attacker is using. In contrast to monitoring, protecting may be very visible to the attacker. Finally, calling a human allows individual discrimination. The IDS can take an initial defensive action immediately while also generating an alert to a human who may take seconds, minutes, or longer to respond.
Intrusion detection systems are not perfect, and mistakes are their biggest problem. Although an IDS might detect an intruder correctly most of the time, it may stumble in two different ways: by raising an alarm for something that is not really an attack (called a false positive, or type I error in the statistical community), or not raising an alarm for a real attack (a false negative, or type II error). Too many false positives means the administrator will be less confident of the IDS's warnings, perhaps leading to a real alarm's being ignored. But false negatives mean that real attacks are passing the IDS without action. We say that the degree of false positives and false negatives represents the sensitivity of the system. Most IDS implementations allow the administrator to tune the system's sensitivity, to strike an acceptable balance between false positives and negatives.
IDS Strengths and Limitations
Intrusion detection systems are evolving products. Research began in the mid-1980s and products had appeared by the mid-1990s. However, this area continues to change as new research influences the design of products.
On the up side, IDSs detect an ever-growing number of serious problems. And as we learn more about problems, we can add their signatures to the IDS model. Thus, over time, IDSs continue to improve. At the same time, they are becoming cheaper and easier to administer.
On the down side, avoiding an IDS is a first priority for successful attackers. An IDS that is not well defended is useless. Fortunately, stealth mode IDSs are difficult even to find on an internal network, let alone to compromise.
IDSs look for known weaknesses, whether through patterns of known attacks or models of normal behavior. Similar IDSs may have identical vulnerabilities, and their selection criteria may miss similar attacks. Knowing how to evade a particular model of IDS is an important piece of intelligence passed within the attacker community. Of course, once manufacturers become aware of a shortcoming in their products, they try to fix them. Fortunately, commercial IDSs are pretty good at identifying attacks.
Another IDS limitation is its sensitivity, which is difficult to measure and adjust. IDSs will never be perfect, so finding the proper balance is critical.
A final limitation is not of IDSs per se, but is one of their use. An IDS does not run itself; someone has to monitor its track record and respond to its alarms. An administrator is foolish to buy and install an IDS and then ignore it.
In general, IDSs are excellent additions to a network's security. Firewalls block traffic to particular ports or addresses; they also constrain certain protocols to limit their impact. But by definition, firewalls have to allow some traffic to enter a protected area. Watching what that traffic actually does inside the protected area is an IDS's job, which it does quite well.