|
|
| Schedule |
| Date: |
Friday, October 9, 2009 |
| Time: |
12:00 — 1:00 pm |
| Location: |
EB 110 |
| Contact: |
(607) 777-4471
|
Abstract
Accurate, robust, convenient conversation between a speaking human and a computer has been a goal of engineers, computer scientists, and linguists for at least the last 30 years as computer power as steadily increased. Although considerable progress has been made in automatic speech recognition (ASR) over these past three decades, for most applications ASR still falls far short of the accuracy of humans. The first part of this presentation will be an overview of why ASR is so difficult, some of the things that have been learned from human speech perception, and some of the tools and algorithms used by ASR researchers. These tools include Hidden Markov Models, Neural Networks, and Gaussian Mixture Models, combined in various ways and “adjusted” for use in automatic speech recognition. Although much current ASR research is focused around higher level sources of knowledge, such as language models, we contend that additional work is needed to determine acoustic features which will enable more accuracy for high quality speech and relatively little degradation in performance for noisy speech. The Discrete Cosine Coefficient Transform/ Discrete Cosine Series (DCTC/DCS) features are proposed as a method to extract spectral/temporal information from sliding “segments” of speech on the order of 500 ms, spaced 8 ms apart. These features thus encode both the short time spectrum, traditionally assumed to contain most of speech information, and the key components of the modulation spectrum, more recently introduced as closely related to the syllabic structure of speech. Although these DCTC/DCS features are quite compact relative to many feature sets used as the front end for ASR systems, the 50 to 100 features used still present challenges as high dimensionality spaces. A nonlinear discriminant analysis will be presented that both reduces dimensionality and increases accuracy. Experimental results for these methods based on the TIMIT and NTIMIT acoustic phonetic database will be given.
Speaker Bio
Stephen A. Zahorian has a BS degree from the University of Rochester, and MS and Ph.D. degrees from Syracuse University, all in electrical engineering. Dr. Zahorian joined the electrical and computer engineering department at Binghamton University in August of 2006 as professor and chairman of the department. He was previously professor and chair of electrical and computer engineering department at Old Dominion University in Norfolk, Virginia. Industry experience includes work as an engineer at RCA Corporation in the Boston area, prior to beginning graduate school. His research and teaching interests are in the areas of signal processing, automatic speech recognition, using computers for biomedical signal processing, and renewable energy. He has obtained over 2 million dollars in total research funding and published over 50 papers in the area of speech signal processing. He and his students have developed a computer-based speech training aid for the hearing impaired. His work has resulted in one patent and one software licensing agreement for multi-media foreign language training. He is a member of the Institute of Electrical and Electronic Engineers (IEEE), the Acoustical Society of America, and the American Society of Engineering Education. He has been active in community outreach activities involving middle and high school students.
Regional Symposium “Graduate Education and Research in Information Security”
GERIS’09, October 26-28, 2009, Binghamton, NY
Purpose and scope:
Infrastructural and information security has become a national priority addressed by regional academic, industrial, and government institutions that often join their efforts in the areas of research and education. Among these institutions are Binghamton University, Syracuse University, Polytechnic University, Utica College, Hamilton College, Lockheed Martin, Assured Information Security, Griffiss Institute, Air Force Research Laboratory at Rome NY. Areas of research include but are not limited to Deception and Counter-Deception, Covert Channels and Steganalysis, Tamper Evidence and Multimedia Forensics, Detection/Mitigation of Attacks on Computer Networks, Watermarking and DRM Systems, and Biologically-Inspired Methods in Computer Security. In the area of education, regional universities offer extensive graduate-level programs focused on the information security that emphasize integration of research in curricula and provide high quality researchers and engineers to the regional and national job market. In addition to local expertise, university researchers have well established collaboration programs with their colleagues nation- and world-wide. Regional Symposium “Graduate Education and Research in Information Security (GERIS’09)” is intended to become a forum for researchers and educators working in this demanding field enabling them to promote the state-of-the art in research, present their findings, share plans and ideas, advance collaborative efforts, assess the requirements of the job market, enhance educational programs, better address the calls of the funding agencies. In addition, the participants will coordinate their efforts in attracting government funding for the NSF’s Integrative Graduate Education and Research Traineeship (IGERT) program and similar collaborative ventures.
The Technical Program of GERIS’09 will include the following four sessions:
- Hardware Security – Session Organizer/Chairman: Prof. Douglas Summerville
- Steganography and Watermarking – Session Organizer/Chairman: Prof. Jessica Fridrich (Binghamton University)
- Computer/Network Security – Session Organizer/Chairman: Prof. Victor Skormin (Binghamton University)
- Graduate Education – Session Organizer/Chairman: Prof. Scott Craver (Binghamton University)
| Schedule |
| Date: |
Friday, September 11, 2009 |
| Time: |
12:00 — 1:00 pm |
| Location: |
EB 110 |
| Contact: |
(607) 777-4471
|
Abstract
Computer networks are continuously subjected to attacks perpetrated by self-replicating malicious software (malware). Modern malware is at least polymorphic and sometimes metamorphic to avoid detection of conventional intrusion detection systems employing binary signatures. However, IDS utilizing behavioral signatures to match malware activity rather than its binary structure are immune to binary morphism. While behavior-based IDSs (BBIDS) have obvious advantages, they potentially suffer from three cross-related problems: run-time efficiency, signature expressiveness and behavioral obfuscation. A signature should be expressive and generic enough to capture a malicious essence of the entire malware family ensuring low false negative and low false positive rates. Behavioral obfuscation is an emerging threat that, given the extensive development of BBIDS, is expected to become a necessary and trivial feature of future malware.
We show approaches to mitigate existing and future challenges of BBIDS. Particularly, we present a novel methodology for behavior specification, generalization and recognition. As a future work, we will discuss approaches for automatic behavior specification extraction and possible application for anomaly based detection.
Bio
Arnur Tokhtabayev is currently a doctoral student in EECE at Binghamton University. He received his Master degree in Electrical Engineering in 2001 from the Kazakh National Technical University, Kazakhstan. His research interests include: intrusion detection in computer networks, local security alarm correlation on network level, intrusion recognition and classification, pro-active defense mechanisms for intrusion prevention.
|
|
