Intrusion Detection and Response
Attack graph analysis identifies critical vulnerability paths and provides strategies for optimal protection of critical network assets. This enables you to make optimal decisions about hardening the network in advance of an attack. But, you must also recognize that because of operational constraints, such as availability of patches and the need for offering mission-critical services, residual vulnerability paths usually remain. But, the knowledge that TVA provides enables you to plan in advance and maintain a proactive security posture even in the face of attacks. For example, TVA attack graphs provide the necessary context for deployment and fine tuning of IDSs, for correlation and prioritization of intrusion alarms, and for attack response.
Intrusion Detection Guidance
Knowledge of vulnerability paths through your network helps you prepare your defenses and your responses. Attacks graphs can guide the optimal deployment and operation of IDSs, which are tailored to your network and its critical assets.
In deploying IDSs, you must decide where to place detection sensors within the network. Traditionally, intrusion detection sensors are placed at network perimeters, with the idea of detecting outside attacks. But, with this deployment, traffic in the internal network is not monitored. If an attacker avoids detection at the perimeter, subsequent attack traffic in the internal network is missed.
On the other hand, deploying sensors everywhere might be cost prohibitive and can overwhelm analysts with floods of alerts. You must strike a balance, where you cover known residual vulnerability paths using the fewest necessary sensors. TVA attack graphs provide this balance.
Consider the attack graph shown in Figure 5-21. Assume that this is the residual attack graph after network hardening measures are applied. So, now the goal is to map this attack graph to the network topology and embed intrusion detection sensors in the network to cover all the vulnerability paths (with the fewest sensors).
Figure 5-21 Residual attack graph is utilized to better determine IDS sensor deployment
Figure 5-22 shows the network topology, overlaid by the attack paths from Figure 5-21. This simplified network diagram illustrates the problem of sensor placement for attack graph coverage. It omits firewalls, which limit connectivity as reflected by the attack graph. Also, the elements labeled router A, router B, and subnet n are abstract network devices capable of monitoring traffic through them (for example, via SPAN ports).
Figure 5-22 Intrusion detection sensor deployment. TVA attack graphs guide the placement of sensors to cover all vulnerability paths while minimizing the number of deployed sensors.
Analysis of the joint topology/attack representation in Figure 5-22 shows that detection sensors placed at router A and router B cover all vulnerability paths with the fewest sensors. An alternative is to place sensors at subnet 1, subnet 4, and subnet 8, which also covers all paths, but requires three (versus two) sensors.
In this network, deploying a sensor at the perimeter alone (router B) misses attack traffic from Servers_1 to Databases. In the opposite extreme, you might decide to deploy sensors at each of the four subnet n devices to catch all potential attack traffic. But, TVA shows that no critical vulnerability paths involve subnet 6, so deploying a sensor there is wasteful, including continually monitoring alerts generated from there. Again, sensors deployed at router A and router B are sufficient to cover all vulnerable paths.
For enterprise networks, performing this kind of analysis requires automation to maximize efficiency. The attack graphs bring together information from various sources over multiple network layers into a concise map. Although the sensor-placement problem itself is hard, a heuristic algorithm scales well and provides near-optimal solutions.29 After sensors are deployed and generate intrusion alarms, you can further leverage attack graphs for alarm correlation and prioritization. This requires mapping alarms to their corresponding elements (exploits) in the residual attack graph. This in turn requires that you represent alarms in a common format, using alarm identifiers that match the identifiers used in the attack graph model.
In this regard, specifications such as Intrusion Detection Message Exchange Format30 (IDMEF) or the ArcSight31 event log format define data formats for information sharing between IDSs and TVA. For example, one implementation option is the IDMEF plug-in32 for Snort.33 This plug-in allows Snort to output alerts in the IDMEF message format. Data exchanges in IDMEF are in XML with the format enforced through a formal schema.
Figure 5-23 shows the structure of an IDMEF alert. The IDMEF model represents alerts in an unambiguous fashion, while explicitly assuming that alert information is heterogeneous. Alerts from different tools might have varying amounts and types of information about an event, which the IDMEF data model accommodates. The critical data is source and target (attacker and victim) network addresses and an alarm identifier that can be mapped to a vulnerability in the TVA model. In IDMEF, these are supported by the Source, Target, and Classification elements, respectively.
Figure 5-23 The IDMEF alert structure provides a standard way to share information between IDSs and TVA.
Attack Prediction and Response
When intrusion alarms are generated, attack graphs provide the necessary context to correlate and prioritize them. First, you can place a high priority on alarms that lie on vulnerability paths through your network. You can prioritize them further based on their graph distance to given critical assets. In other words, events that are close to critical assets (in terms of next attack steps) are given a higher priority compared to resources buried deep in the infrastructure.
This kind of attack graph analysis is highly precise and takes all relevant facts into account. You determine not only whether a host is vulnerable to a given attack, but whether the attacker can traverse through firewalls to reach the host's vulnerable port and whether that attack can lead to subsequent network compromise. Thus, your prioritization also serves as an advanced form of false-alarm reduction, restricting alarms along critical paths.
It is important to model network vulnerability because multistep alarm correlation do not take real network vulnerabilities into account often.34 Precomputing vulnerability-based attack graphs in advance of an attack has the additional advantage of rapid correlation, which means that it's faster than an IDS can generate them.35,36
Furthermore, the predictive capabilities of attack graphs enable you to correlate intrusion alarms based on attack causality. A set of seemingly isolated events might in fact be shown as multiple steps of incremental network penetration. Also, the context provided by these attack graphs enables you to predict potentially missed events (false negatives), which helps mitigate inaccuracies in your defense posture.37
To illustrate some of these ideas, consider Figure 5-24. This is the same residual attack graph shown in Figure 5-21, but with relevant protection domains expanded to show additional details. This attack graph provides considerable insight for correlating and prioritizing any alarms generated for this network and for responding to these potential attacks.
Figure 5-24 Attack prediction and response allows analysts to better determine risk
For example, suppose an alarm is raised for an attack between two machines in the DMZ (say, from DMZ_1 to DMZ_2). From just a single alarm in the DMZ, you might wait before responding. On the other hand, if an alarm is raised from Internet into DMZ, followed by an alarm within the DMZ, it is a stronger indicator that the attack might be a real security breach. Remember that false alarms are common with intrusion detection, and erroneously blocking traffic in response to false alarms is a denial of service.
From an alarm within the DMZ, another approach might be to block traffic from DMZ_3 to DB_1 and DB_2. Because of the possibility of denial of service, such an action is not usually taken. But you can limit the blocking to the vulnerable ports on DB_1 and DB_2 only, specifically from DMZ_3, so that any nonvulnerable services on those machines can remain unblocked. You might then keep traffic from DMZ_3 into Servers_1 machines unblocked, because those machines are one less attack step (three steps) from critical machine DB_4. In other words, you can wait to see if an alarm is raised from the DMZ into Servers_1, at which point you block the vulnerable paths from Servers_1 to Databases.
An even more aggressive response to an alarm within the DMZ is to block outgoing traffic from the DMZ to vulnerable services in Servers_1 and Databases. Again, there is the potential for denial of service, but you still limit your response to vulnerable connectivity. Without attack graph analysis, the only response to a serious attack is to block all traffic from the DMZ, not just vulnerable connectivity. Furthermore, you can surmise that an alarm in the DMZ is follow-on from a missed intrusion from the Internet into the DMZ. This can guide further investigation into traffic logs into the DMZ looking for missed attacks, especially against the four vulnerable paths into the DMZ.
If an attack was detected within Servers_1 (such as from Server_1 to Server_2), a similar set of responses is indicated. As a precaution, you could block traffic from Server_3 to vulnerable ports on DB_1 and DB_2. But, blocking traffic from Server_3 into the DMZ is less indicated because it leads away from the critical Databases domain. Similarly, any alerts from Server_3 into the DMZ are lower priority, especially if they are not against vulnerable DMZ services.
Thus, provides a range of reasonable responses, ranked by severity or actual likelihood of attack. Here, severity is in terms of lying on critical vulnerability paths, especially close to critical assets, and its likelihood increases by causal correlation of alerts. Multiple options are available that enable you to fine tune responses as potential attacks unfold, based on proactive response plans.