Eric Breimer

Intrusion detection: a bioinformatics approach

We address the problem of detecting masquerading, a security attack in which an intruder assumes the identity of a legitimate user. Many approaches based on hidden Markov models and various forms of finite state automata have been proposed to solve this problem. The novelty of our approach results from the application of techniques used in bioinformatics for a pair-wise sequence alignment to compare the monitored session with past user behavior. Our algorithm uses a semiglobal alignment and a unique scoring system to measure similarity between a sequence of commands produced by a potential intruder and the user signature, which is a sequence of commands collected from a legitimate user. We tested this algorithm on the standard intrusion data collection set. As discussed, the results of the test showed that the described algorithm yields a promising combination of intrusion detection rate and false positive rate, when compared to published intrusion detection algorithms.

A task-based architecture for application-aware adjuncts

Users of complex applications need advice, assistance, and feedback while they work. We are experimenting with “adjunct” user agents that are aware of the history of interaction surrounding the accomplishment of a task. This paper describes an architectural framework for constructing these agents. Using this framework, we have implemented a critiquing system that can give task-oriented critiques to trainees while they use operating system tools and software applications. Our approach is generic, widely applicable, and works directly with off-the-shelf software packages.

Learning Significant Alignments: An Alternative to Normalized Local Alignment

We describe a supervised learning approach to resolve difficulties in finding biologically significant local alignments. It was noticed that the O(n2) algorithm by Smith-Waterman, the prevalent tool for computing local sequence alignment, often outputs long, meaningless alignments while ignoring shorter, biologically significant ones. Arslan et. al. proposed an O(n2 log n) algorithm which outputs a normalized local alignment that maximizes the degree of similarity rather than the total similarity score. Given a properly selected normalization parameter, the algorithm can discover significant alignments that would be missed by the Smith-Waterman algorithm. Unfortunately, determining a proper normalization parameter requires repeated executions with different parameter values and expert feedback to determine the usefulness of the alignments. We propose a learning approach that uses existing biologically significant alignments to learn parameters for intelligently processing sub-optimal Smith-Waterman alignments. Our algorithm runs in O(n2) time and can discover biologically significant alignments without requiring expert feedback to produce meaningful results.

On the Height of a Random Set of Points in a

We investigate, through numerical experiments, the asymptotic behavior of the length Hd(n) of a maximal chain (longest totally ordered subset) of a set of n points drawn from a uniform distribution on the d-dimensional unit cube V D = [0, 1]d. For d ≥ 2, it is known that cd(n) = Hd(n)/n1/d converges in probability to a constant Cd < e, with Iim d→∞ Cd = e. For d = 2, the problem has been extensively studied, and it is known that C2 = 2; Cd is not currently known for any d ≥ 3. Straightforward Monte Carlo simulations to obtain Cd have already been proposed, and shown to be beyond the scope of current computational resources. In this paper, we present a computational approach which yields feasible experiments that lead to estimates for Cd. We prove that Hd(n) can be estimated by considering only those chains close to the diagonal of the cube. A new conjecture regarding the asymptotic behavior of cd(n) leads to even more efficient experiments. We present experimental support for our conjecture, and the new estimates of Cd obtained from our experiments, for d ∈ {3,4,S,6}.

d

We develop computationally feasible algorithms to numerically investigate the asymptotic behavior of the length Hd(n) of a maximal chain (longest totally ordered subset) of a set of n points drawn from a uniform distribution on the d-dimensional unit cube Vd = [0, 1]d. For d ≥ 2, it is known that cd(n) = Hd(n)/n1/d converges in probability to a constant cd < e, with limd→∞ cd = e. For d = 2, the problem has been extensively studied, and, it is known that c2 = 2. Monte Carlo simulations coupled with the standard dynamic programming algorithm for obtaining the length of a maximal chain do not yield computationally feasible experiments. We show that Hd(n) can be estimated by considering only the chains that are close to the diagonal of the cube and develop efficient algorithms for obtaining the maximal chain in this region of the cube. We use the improved algorithm together with a linearity conjecture regarding the asymptotic behavior of cd(n) to obtain even faster convergence to cd. We present experimental simulations to demonstrate our results and produce new estimates of cd for d ∈ {3,...,6}.