Guessing Secrets

Applying mathematics to the efficient delivery of Internet content

Even on uneventful days, traffic on the Internet can sometimes stutter to a crawl. It gets much worse when millions of people go online at the same time to view the latest images from a Mars expedition, download a trailer for an upcoming Star Wars movie, or take in a titillating fashion show. The mushrooming demand on such days can rapidly clog this worldwide web of computer networks, causing horrendous delays and outages. In other words, access to Web sites melts down just as things get interesting.

GLOBAL CONNECTIONS. Akamai’s network operations center. Akamai

“We have to use the Internet the way it is, bugs and all,” says mathematician Tom Leighton of the Massachusetts Institute of Technology, one of the founders of Akamai Technologies in Cambridge, Mass. Originally designed several decades ago to handle communication among researchers at a handful of laboratories, the system that’s now the Internet can falter in the face of massive, global migrations of digital data.

Since 1999, Akamai has offered to highly popular Web sites ways to ease congestion. The company redistributes text, images, and movies through its own computer network, which is independent of but connected to the Internet.

Akamai’s network takes advantage of sophisticated mathematical methods to determine which of the company’s worldwide collection of more than 14,000 computers should store a Web site’s content so that it can unfailingly get to users in the shortest possible time. Akamai’s customers include the Centers for Disease Control, retailer Victoria’s Secret, and the search engine Yahoo.

Leighton, graduate student Danny Lewin, and several colleagues developed the proprietary algorithms that govern the way Akamai manages and redistributes information. (Lewin died on Sept. 11, 2001, aboard one of the planes that struck the World Trade Center.)

The data-management tools developed by Akamai to keep Web sites operating efficiently solve about 99 percent of the problem of Internet-style traffic jams, Leighton says. To better that performance, Akamai researchers must tackle various quirks in the rules that govern Internet communication.

One recent effort to improve network performance has mathematicians taking a fresh look at the familiar game of 20 questions. In this game, a player tries to identify a secret object by asking a sequence of questions–traditionally beginning with “Animal? Vegetable? Mineral?”– that can be answered by yes or no.

In the Internet variant of the game, the secret is a sequence of 32 binary digits representing a computer’s Internet protocol (IP) address. In this game, however, there are always two or more secret solutions, and the responder supplies a truthful answer to a given question without specifying for which secret the answer is true.

“The research issues are: How much can you learn, and how quickly can you do it?” says Ronald L. Graham of the University of California, San Diego. “It’s a fascinating problem that has surprising links to all sorts of mathematics.”

Internet match game

Roaming the Internet’s World Wide Web appears effortless. To join the hordes eager for a glimpse of a newly imaged Martian rock, for instance, you simply fire up your Web browser and type in a Web-page identifier–the so-called uniform resource locator, or URL. If the image is available only at a single computer that’s sitting in, say, a NASA laboratory, every such request must find its way to that one Internet address. The frequent result is a mess of jammed communication lines, an overwhelmed Web site, and frustrated armchair explorers.

A significant part of Akamai’s strategy is to make high-demand content available at multiple computers throughout the world. The next step is to match Web-page requests with the appropriate Akamai servers–in effect, shortening the paths traversed by requests and data throughout the Internet labyrinth. For the Akamai system to decide which server should deliver the requested content, it’s desirable to quickly and accurately pinpoint the geographic location of a user’s computer.

Every computer connected to the Internet has its own numerical address, but there’s a burdensome complication in the way the system operates. When a browser makes a request, the message goes to a network computer known as a nameserver, which looks up the target Web site’s numerical Internet address and passes the message on to the relevant server. For example, any request for a Science News Online Web page (at https://www.sciencenews.org/) must be sent to a computer with the IP address 216.167.111.80.

If the message is directed to a Web site for an Akamai customer (Science News Online is not one), the Akamai system receives the message and must decide which of its servers should provide the content. However, all that system sees initially is the Internet address of the nameserver, not of the client. The nameserver’s address doesn’t reliably indicate locations of the computers it serves.

So, it would be helpful for Akamai to know which nameservers handle which clients. In a quirk of the protocols governing Internet communication, the address of a nameserver never appears together with the address of any one of its clients. However, the nameserver can provide some additional information, which might be used to deduce the client address.

In effect, the strings of digits representing clients’ Internet addresses need to be guessed using a step-by-step process analogous to the game of 20 questions.

A home computer with only one link to the Internet would probably have only one address. That address would be relatively easy to determine: A questioner could ask about each digit in turn. Many computers, especially in large corporations, have two or more addresses for added security and to distribute the activity.

The game variant that proved of particular interest to mathematicians involves a client computer with two Internet addresses–in other words, two secrets–either of which may be in use at any given moment. In this case, how much can you hope to learn about both secrets (IP addresses) from yes-or-no answers? It would be as if in 20 questions, the answerer had two objects in mind and answered “yes” when it was appropriate for either object.

Leighton, Graham, and San Diego mathematician Fan Chung tackled this problem.

Their findings appear online in the Electronic Journal of Combinatorics (http://www.combinatorics.org/Volume_8/Abstracts/v8i1r13.html).

They deliberately chose the most difficult cases, and the news is discouraging. The mathematicians say that there’s no way to guarantee that you can learn both secrets from a “malevolent but truthful” adversary. It’s possible, for example, that all replies apply to just one of the two secrets.

You would then learn nothing about the other.

It gets worse. The mathematical analysis shows that it isn’t possible to guarantee that you can learn even one element of a secret–say, one binary digit of an Internet address–let alone the entire secret.

In this simplified, hypothetical example, suppose you’ve narrowed down the possibilities for the final digits in the two secret Web addresses to two of three four-digit strings: 1 0 0 1, 1 1 0 1, or 0 0 1 1. You ask whether the second digit is 1. The adversary can stymie your efforts to learn anything about the secret pair by choosing the answer that applies to the majority of the three strings. In this case, because 0 comes up twice as the second element and 1 only once, the answer would be “no.” Similarly, if you ask whether the sum of the digits is odd, the respondent would answer “no”. Based on majority answers, no matter which questions you ask (or how many), you can’t narrow down the choices to a single pair.

The mathematicians obtained additional insights by expressing the problem in terms of graph theory, representing possible secrets as pairs of points joined by lines to form a network (see below). Yes-or-no answers to questions eliminate pairs from consideration, in the end leaving either a single line identifying the two secrets or certain configurations of lines and points that can’t be resolved further.

The same problem can be extended from a game with two secret answers to a game with three. “There’s a big jump in complexity from two to three,” Graham observes. However, “we can enumerate the many cases where definitive answers can’t be obtained.”

At the same time, there is some good news. Mathematicians have developed very quick ways for getting to the answer or, in the worst cases, narrowing it down to a few possibilities.

Intriguing puzzles

Although it originated in an Internet context, the mathematical problem that underlies guessing secrets has taken on a life of its own as an intriguing puzzle and as a practical concern in computer science. Graham, Chung, and others have recently explored what sorts of questions to ask to arrive at the final answers–in cases where it can be done–in the fewest steps.

“You need to ask good questions–ones that can eliminate many possibilities,” Chung says. For example, when there are two or more secrets expressed as binary digits, more possibilities are ruled out when you ask whether the sum of the secret digits is an even number than when you ask whether the first digit is 1.

Computer scientists have discovered connections between the problem of guessing secrets and various topics in computer science, such as separating systems into smaller units, diagnosing technical problems, protecting data from unauthorized reproduction, and authenticating ownership claims. Graham himself was surprised to find that his current effort is related to research he originally did in the 1960s on the performance of electrical circuits.

When MIT computer scientist Madhu Sudan heard about the Chung-Graham-Leighton work, he noticed a link to a task known as list decoding, which concerns errors that may occur during the transmission of digital data.

“When the number of errors is guaranteed to be small, and one puts in sufficiently large amounts of redundancy in encoding one’s transmission, then it is possible to pin down the transmitted message,” Sudan says. “When the number of errors is somewhat larger, it may not be possible to pin down the transmitted message.”

It may be possible, however, to isolate a small set, or list, of messages that include the transmitted message. Computer scientists call this process “list decoding.”

By establishing a connection between the problem of guessing secrets and list decoding, Sudan and his coworkers developed procedures for recovering secrets more efficiently than by using methods originally proposed by Chung and her colleagues.

That’s only the beginning. “There are numerous questions about guessing secrets that remain unanswered,” Chung remarks. For instance, no one has yet looked at what happens in the case of guessing two secrets if some of the answers to queries aren’t true.

Additional applications of this type of research may also be on the horizon. The notion of guessing secrets shows promise in the context of making digital information more secure, Sudan suggests.

Akamai’s networks haven’t yet felt the impact of this research. “We have not used this work for the original applied problem because we are currently following a different . . . approach,” Leighton says. “But we might use some of the stuff over the course of the next year.”

He adds, “It certainly was interesting to encounter such a deep and rich mathematical problem when working on a seemingly applied problem.”

That’s not an unusual occurrence at Akamai.

“We have a lot of people here with strong mathematical backgrounds,” Leighton notes, “so it is natural for us to recognize interesting mathematical problems that lurk beneath more applied problems involving the Internet.”


A Graphic Route for Disclosing Secrets

Guessing a pair of secrets can be visualized in terms of an array of points and lines, an array that mathematicians describe as a graph. Each point represents a potential answer, and a line links each pair. The player zeroes in on the pair of secrets by asking yes-or-no questions. Each answer may eliminate certain lines from further consideration. In the best case, the result is two points connected by a line. In the worst possible case, however, a player may be left with a triangle of three potential secrets or a star made up of several possible secrets. In neither of these instances is it possible for the player to identify both secrets.–I.P.


More Stories from Science News on Math