5.5 Extrusion Detection
Extrusion detection is a specialized form of IDS. It’s aimed at one particular form of harm: someone trying to steal your data and export it. The trick is picking up the outbound data transfer. There are two challenges: picking out the right data, amidst all of the legitimate (or at least normal) traffic, and distinguishing authorized from unauthorized transfers. This latter isn’t trivial; uploading a chip design to a foundry can be the normal way of doing business, while sending it to the Andromedans’ web server most likely is not. Extrusion detection has one principal advantage over many other types of security systems: it can cope with rogue insiders.
There are a number of ways to perform extrusion detection. One of the simplest is the honeypot: create fake files that will attract the attention of a spy, commercial or governmental, and wait for someone to grab one. This has been done a number of times, most famously by Stoll [1988; 1989] in the “Wily Hacker” incident. (Bill Cheswick and I subtitled Firewalls “Repelling the Wily Hacker” in homage to Cliff, and we used that phrase with his permission.) Briefly, he discovered intruders in a University of California computer system and traced them to Germany. To allow enough time for technicians to trace the attackers’ phone calls—this was in the days of dial-up modems—he created fake documents about the Strategic Defense Initiative, a missile defense system, and waited for someone to look at them. His trap was successful; the attackers’ response included sufficient indicia of espionage that he notified the FBI.
In the more usual scenario, you don’t know whether you’ve been penetrated. Accordingly, the proper honeypot strategy requires a wide range of believable-seeming decoys. Exactly what decoys you should create depends on your system; you want something that resembles normal items on that system. Bowen at al. [Bowen 2011; Bowen et al. 2009] describe a number of types of decoys—“honey documents”—including fake logins for banking web sites; they also define criteria for decoy generation. One notable aspect to their work was the use of web bugs, normally used by advertisers to track consumers on the web. With a web bug, opening the document causes an attempt to fetch a file (typically a 1 × 1–pixel transparent GIF) from a monitored HTTP server.
Naturally, a sufficiently knowledgeable attacker can dodge many decoys. Someone who suspects web bugs, for example, would simply read exfiltrated documents on an offline computer. Indeed, intelligence agencies’ classified networks are generally disconnected from the outside world [R. A. Clarke and Knake 2010], so no strategy that relies on active documents can succeed. A different approach is needed, one based on IDS technology.
The big advantage of honey documents is that they’re transport independent. That is, no matter how the files are exfiltrated, the trap can be sprung on any network connected machine used to view them. Even printing them out first doesn’t help; the monitors will detect the documents being opened inside the enterprise.
You may be able to detect network based exfiltration while the documents are being transported if your network is configured in a firewall-friendly manner, that is, if Property 1 holds. An extrusion detection module can be installed at the firewall; it can then attempt to detect misbehavior from amidst the noise of routine Internet traffic. As with IDSs, one can approach this from a signature or anomaly detection perspective. Signature detection can look for certain documents or perhaps markings—should the strings “Company Confidential” or “Top Secret UMBRA” appear in outbound mail?—or anomaly detection. Anomaly detection might be as simple as volume—does this person or IP address normally send so much data? Does someone in that organization normally send so much?—or it may be based on the statistical characteristics of the outbound data.
There are some interesting wrinkles here that makes extrusion detection harder in some ways than firewalls or intrusion detection. For one thing, someone exporting information is freer to use encryption because he or she can control both ends of the channel. By contrast, an attacker breaking in to an organization can only encrypt traffic if the vulnerability being exploited uses encryption. Of course, once the penetration is complete, the attacker can install any sort of back door desired, including encrypted ones. Even if encryption is used, the defenders aren’t helpless. While crypto does hide the precise content being sent, it can’t hide the volume; more importantly, encrypted data has a very unique flat byte distribution; this, too, is anomalous if from a source or to a destination that does not normally receive such. An example is shown in Figure 5.1, which compares the byte frequency distribution of a JPG file and an encrypted version of the same file.
Figure 5.1: A byte frequency plot for a JPG file and an encrypted version of the same file. (The x = 0 value was omitted, since the frequency of 0 in the JPG file was so high that it obscured the other differences.) The horizontal line is , what a perfectly even distribution would be. Note that JPG and other highly compressed formats like MP3 are harder to distinguish from random data; encrypted files in general do (and should) look very random. In a test of 1024-byte blocks, Oprea et al. found that less than 2% had an entropy that appeared close to that of random data .
In some situations, the firewall-like topology need not be of the user network, but rather of the data-providing network. Imagine a data center with a web front end serving thousands of clients around the world. There is no feasible perimeter around the users; however, all sensitive data is flowing out through that web server. Proper log files and analysis of them (see Section 16.3) can also be a form of extrusion detection. Indeed, the lack of such analysis has been claimed to be at fault in the WikiLeaks diplomatic cables case; see the box on p. 73. Some organizations already track all outbound traffic. In the United States, some financial sector companies are required to log and retain certain communications [FINRA 2010; FINRA 2011]; these records can be analyzed for improper exports. Of course, the rest of the perimeter needs to be secured to preclude other forms of export.
One can view intrusion and extrusion detection systems as in some sense the dual of firewalls. The latter attempts to prevent trouble; the former attempt to detect it. Firewalls are primarily concerned with what the communications endpoints and protocols are; detection systems are more concerned with the contents. Firewalls are generally centralized; detection systems function better if decentralized. Both yield benefits; both have their limitations.