Colloquium

I will describe the 2nd generation Onion Routing* system, Tor,which is freely available from http://freehaven.net/tor/. I will give backround on anonymous communication in general and on related systems. If there is time, I will also mention some of the theory in defining and measuring anonymity as well as the economic and intellectual property issues that have arisen in developing and deploying such an infrastructure.

*Onion Routing is an infrastructure for private communication over a public network, such as the Internet. It provides anonymous connections that are resistant to both eavesdropping and traffic analysis. Thus it hides not only the data being sent, but who is talking to whom. Onion Routing provides real-time, bidirectional connections that can be used anywhere a socket connection can be used. Applications, such as web browsing, e-mail, chat, and remote login require no modification to use Onion Routing.

In order to enable computers to understand and use natural language, a massive amount of knowledge, linguistic and otherwise, must be amassed. As a result, much recent research has focused on creating systems that automatically learn high-quality information about language, and about the world, directly from the statistics of unprocessed or minimally-processed language samples alone. As examples, I will focus on two lines of work. The first uses information-theoretic distributional clustering methods trained on large language samples to induce probabilistic models of linguistic co-occurrences. The second utilizes multiple-sequence-alignment algorithms, commonly employed in computational biology, to learn how to generate English versions of computer-generated proofs, creating texts whose quality rivals that of hand-crafted systems.

Portions of this talk are based on joint work with Regina Barzilay and with Fernando Pereira.

Models for the processes by which ideas and influence propagate through a social network have been studied in a number of domains,including the diffusion of scientific and technological innovations,the sudden and widespread adoption of various strategies in game-theoretic settings, and the effects of ``word of mouth'' in the promotion of new products. Emerging on-line media such as Weblogs offer a compelling new setting in which to study such phenomena, since they are characterized by frequent ``topic bursts'' in which a virtual discussion spreads rapidly through the network of individuals who author on-line content.

A natural way to identify highly influential individuals in such network settings is to look for collections of nodes with the power to trigger large outbreaks of cascading behavior. In other words, if we can try to convince a subset of individuals to adopt a new behavior (e.g. to try a new product or take part in an on-line discussion),and the goal is to cause a large cascade of further adoptions, which set of individuals should we target? In this talk, we will consider such problems in several of the most widely studied models in social network analysis, and describe algorithms (obtained in joint work with David Kempe and Eva Tardos) with provable approximation guarantees for the problem of selecting the most influential set of nodes. The development of these algorithms indicates how influential sets must include members from diverse parts of the network, forcing heuristics for this problem to deal, at least implicitly, with the clustering of the underlying opulation. We conclude by discussing the intriguing relationship between models of influence propagation and the problem of identifying ``bursty'' topics, which is emerging as an important issue for searching on-line information resources.

The Internet is a network of competing networks. Independent network operators have access to the proprietary details of their own networks, but neither researchers nor network operators have had access to the detailed, global picture of the Internet needed to find and correct network vulnerabilities and evaluate new applications and protocols. My thesis work demonstrates that detailed topological and routing information regarding the global Internet and its constituent networks is within the reach of these communities. That is, that reverse-engineering the Internet can be made practical through innovative measurement and inference techniques. In this talk, I present my Rocketfuel system which efficiently maps networks in the Internet using only externally available information. To recover a reasonably complete network graph, Rocketfuel uses hundreds of traceroute servers as vantage points to collect paths through the network. For efficiency, Rocketfuel uses global routing information and prior measurements to guide further measurement. To give the maps structure, Rocketfuel uses the information encoded in router names to assign each router to a geographic location and uses the rest of the network graph to determine the role of each router in the network. Topology is only part of the picture; the rules that govern how packets are directed across the topology are equally important but impossible to measure directly. I will describe how routing can be inferred from paths not taken. Inferred routing allows us to summarize and predict which path a packet will traverse, as well as to find interesting configuration decisions. Understanding global Internet topology and routing allows operators to predict the effects of change and allows researchers to identify problems and demonstrate the effectiveness of their solutions using realistic Internet topologies.

Most classification algorithms were originally designed for fixed-size vectors. However, important learning problems in natural language processing require the analysis of variable-length sequences and more generally distributions over variable-length sequences.

Rational kernels are a new family of similarity measures over variable-length sequences and their distributions. Many similarity measures commonly used in computational biology, such as the edit distance, the convolution kernels of Haussler, and other string kernels, are shown to be special cases of rational kernels.

This talk will describe general and efficient methods for computing rational kernels, and discuss some important convergence and closure properties. It will also report the results of experiments illustrating the successful use of rational kernels for several difficult prediction problems in text and speech processing.

[Joint work with Corinna Cortes and Patrick Haffner]

One of the main reasons for the success of the internet is its service model that emphasizes flexibility. While this freedom enabled the widespread deployment of the applications popular today such as email and Web, it has also greatly complicated the task of administering these networks. To understand how a network is being used, or whether it is being abused, an administrator must inspect the flow of packets and "infer" the intent of users and applications. Existing measurement solutions either lack the necessary detail, do not scale up to the speeds of today's networks or are not flexible enough to keep up to the speeds of today's networks or are not flexible enough to keep up with the ever changing traffic mix. I will present two approaches to improve the state of the art.

My first approach is to develop fast and accurate algorithmic building blocks that allow routers to collect better measurement data. For example, it is often necessary to identify large flows of traffic, the "heavy hitters". I will present multistage filters which quickly and scalably identify heavy-hitters. A second useful building block scalably estimates the number of active flows or IP addresses using a family bitmap algorithms. I will show theoretical and experimental evaluations of the effectiveness of these building blocks.

My second approach is to improve the flexibility of offline analysis through a new method of traffic characterization. The conventional approach is a static analysis specialized to capture flows, applications, or network-to-network traffic matrices. By contrast, my analysis dynamically and automatically produces hybrid traffic definitions that match the underlying usage. I will describe a publicly available tool called AutoFocus that I built to implement this analysis, and its use on various production networks to infer such varied phenomena as new worms, denial of service attacks, routing changes, and traffic periodicities.

The frameworks of game theory and mechanism design have exerted significant influence on formal models of multiagent systems and electronic markets by providing tools for designing and analyzing systems in order to guarantee certain desirable outcomes. However, many game-theoretic models assume idealized rational decision makers interacting in prescribed ways. In particular, the models often ignore the fact that in many systems the agents are not fully rational, but instead have constraints on their computational capabilities which inhibit them from acting as fully rational agents would. This creates a potentially hazardous gap in game theory and automated negotiation as agents may not be motivated to behave in the expected way, leading to potentially undesirable outcomes.

In this talk I will present work aimed at bridging this gap. By explicitly incorporating the deliberation (computational and information gathering) actions of agents into their strategies, we are able to provide a theory of interaction for self-interested computationally-bounded agents. I will describe a new game-theoretic solution concept, the deliberation equilibrium and use this concept to analyze several classic and commonly used auctions. In particular, I will illustrate the complex strategic behavior which arises when agents have deliberation limitations, which has been overlooked by traditional analysis. I will conclude by discussing mechanism design principles for multiagent systems and electronic markets with resource-bounded agents.