When systems are infected with malware, their communication patterns usually change in a detectable fashion. Consider this example:
Client machines usually talk with servers. Servers very seldom initiate a connection back to clients, except for occasional services like File Transfer Protocol (FTP) not used in passive mode. Also, clients almost never communicate with other clients, and servers have only a little communication with other servers. Hence you have a nice pattern that automated tools can check for.
When a worm or bot infection occurs, there is often a huge uptick in client-to-client session initiation. As you point out in your question, there might be a major rise in the bandwidth consumption of one or more infected machines. There also may be a hike in the number of connection initiation attempts. Each of these measurements is helpful and can be detected by various NBAD products. Network-based intrusion prevention systems, security information management (SIM) products, some intrusion detection systems, as well as distributed denial of service (DDoS) monitoring products all offer such capabilities.
Beyond these products, there are large-scale, Internet-based projects that look for network anomalies. One of the most prominent is the DShield project, administered by the SANS Internet Storm Center. This project has over 45,000 volunteer-operated sensors distributed around the Internet. The sensors gather data, make it anonymous and send it to collectors. Software and people then analyze the resulting information, which includes communicating sessions and the ports they use. The top 10 worldwide rising ports, as well as various unusual session activity, are plotted and updated every day on the DShield Web site.
This was first published in June 2007