Network-based Threat Detection: Looking for IndicatorsBy Mike Rothman
Now that RSAC is behind us, it’s time to get back to our research agenda. So we pick up Network-based Threat Detection where we left off. In that first post, we made the case that math and context are the keys to detecting attacks from network activity, given that we cannot totally prevent endpoint compromise. Attackers always leave a trail on the network.
So we need to collect and analyze network telemetry to determine whether the communication between devices and the content of communications are legitimate, or warrant additional investigation. Modern malware relies heavily on the network to initiate the connection between the device and the controller, download attacks, perform automated beaconing, etc. Fortunately these activities show a deterministic pattern, which enables you to pinpoint malicious activity and identify compromised systems.
Attackers bet they will be able to obscure their communications within the tens of billions of legitimate packets traversing enterprise networks on any given day, and on defenders’ general lack of sophistication preventing them from identifying the giveaway patterns. But if you can identify the patterns, you have an opportunity to detect attacks.
Command and Control
Command and Control (C&C) traffic is communication between compromised devices and botnet controllers. Once the device executes malware (by whatever means) and the dropper is installed, the device searches for its controller to receive further instructions. There are two main ways to identify C&C activity: traffic destination and communications patterns between.
The industry has been using IP reputation for years to identify malicious destinations on the Internet. Security researchers evaluate each IP address and determine whether it is ‘good’ or ‘bad’ based on activity they observe across a massive network of sensors. IP reputation turns out to be a pretty good indicator that an address has been used for malicious activity at some point. Traffic to known-bad destinations is definitely worth checking out, and perhaps even blocking. But malicious IP addresses (and even domains) are not active for long, as attackers cycle through addresses and domains frequently.
Attackers also use legitimate sites as C&C nodes, which can leave innocent (but compromised) sites with a bad reputation. So the downside to blocking traffic to sites with bad reputation is the risk of irritating users who want to use the legitimate site. Our research shows increasing comfort with blocking sites because the great majority of addresses with bad reputations have legitimately earned it.
Keep in mind that IP reputation is not sufficient to identify all the C&C traffic on your network – many malicious sites don’t show up on IP reputation lists. So next look for other indications of malicious activity on the network, which depends on how compromised devices find their controllers.
With the increasing use of domain generating algorithms (DGA), malware doesn’t need to be hard-coded with specific domains or IP addresses – instead it cycles through a set of domains according to its DGA, searching for a dynamically addressed C&C controller; the addresses cycle daily. This provides tremendous flexibility for attackers to ensure newly-compromised devices can establish contact, despite frequent domain takedowns and C&C interruptions. But these algorithms look for controllers in a predictable pattern, making frequent DNS calls in specific patterns. So DNS traffic analysis has become critical for identification of C&C traffic, along with monitoring packet streams.
Identifying C&C traffic before the compromised device becomes a full-fledged member of the botnet is optimal. But if you miss, once the device is part of the botnet you can look for indications that it is being used as part of an attack chain. You do this by looking for outliers: devices acting atypically.
Does this sound familiar? It should – anomaly detection has been used to find attackers for over a decade, typically using Netflow. You profile normal traffic patterns for users on your network (source/destination/protocol), and then look for situations where traffic varies outside your baseline and exceeds tolerances.
Network-based anomaly detection was reasonably effective, but as adversaries got more sophisticated detection needed to dig more deeply into traffic. Deep packet inspection and better analytics enabled detection offerings to apply context to traffic. Attack traffic tends to occur in a few cycles:
- Command and Control: As described above, devices communicate with botnet controllers to join the botnet.
- Reconnaissance: After compromising the device and gaining access via the botnet, attackers communicate with internal devices to map the network and determine the most efficient path to their target.
- Lateral Movement: Once the best path to the target is identified, attackers systematically move through your network to approach their intended target, by compromising additional devices.
- Exfiltration: Once the target device is compromised, the attacker needs to move the data from the target device, outside the network. This can be done using tunnels, staging servers, and other means to obfuscate activity.
Each of these cycles includes patterns you can look for to identify potential attacks. But this still isn’t a smoking gun – at some point you will need to apply additional context to understand intent. Analyzing content in the communication stream is the next step in identifying attacks.
One way to glean more context for network traffic is to understand what is being moved. With deep packet inspection and session reassembly, you can perform file-based analysis on content as well. Then you can compare against baselines to look for anomalies in the movement of content within your network as well.
- File size: For example, if a user moved 2gb of traffic over a 24 hour period, when they normally move no more than 100mb, that should trigger an alert. Perhaps it’s nothing, but it should be investigated.
- Time of day: Similarly, if a user doesn’t normally work in the middle of the night, but does so two days in a row by themselves, that could indicate malicious activity. Of course it might be just a big project, but it bears investigation.
- Simple DLP: You can fingerprint files to look for sensitive content, or regular expressions which match account numbers or other protected data. Of course that isn’t full DLP-style classification and analysis. But it could flag something malicious without the overhead of full DLP.
Content analysis won’t to provide a smoking gun either. But with network traffic detection as discussed above, it provides more context to start to discern intent. This context helps explain some behavior that would otherwise be flagged as anomalous, to reduce false positives.
Malware crossing the perimeter does not necessarily mean it executed on any devices. That is a weakness of network-based sandboxes, which just look at and alert on files coming into the network. Those devices fire an alert whenever they see malware, even if the target device is totally protected from the attack. One way to further identify real attacks is to integrate endpoint telemetry into analysis, to verify and validate what actually happened. So we increasingly see a drive for network-based detection coordinated with endpoint detection.
That doesn’t mean you don’t want to know malware entered the network, but you need some way to prioritize whether or not it needs to be dealt with right now. Which brings up the much larger issue of prioritization, and knowing which potential attack to deal with first. It comes down to understanding what presents the most clear and present danger (risk) to your environment, which we’ll tackle in our next post.