Rowan professor helps computers keep pace with evolving challengesMay 23, 2013
As increasingly crafty spammers, malware writers and other malicious programmers persistently devise new tactics to outwit sophisticated computer systems, a Rowan University College of Engineering professor is researching tools that may help computers stay one step ahead of these risks.
Dr. Robi Polikar, professor and chair of Electrical and Computer Engineering at the Rowan University College of Engineering, hopes to develop automated algorithms — mathematical procedures differentiating different classes of data — that may be used in recognizing and detecting such changing threats and give computers an edge in defeating these adversaries. The primary novelty of the developed algorithm is its ability to learn from streams of unstructured data of changing characteristics and adapt and respond to evolving challenges associated with such data.
Analyzing variable data
Experts know how to develop algorithms that analyze data displaying characteristics that do not change — known as stationary data — and make decisions based on previous and historic patterns, a concept now well established in machine learning and computational intelligence. For example, when a computer analyzes biological data from patients with a particular disorder, another of Polikar’s areas of research, the known disease characteristics for a given cohort do not change over time in most cases, facilitating diagnosis.
“The particular problem we are trying to solve in more fundamental machine learning research is: Can you make this decision based on large volumes of data if the characteristics of the data change over time?” Polikar said. “For example, as we get better in detecting spam, the spammers find new ways to spam people.” Or the characteristics of the data in malware attacks or financial fraud change over time as well. These are referred to as nonstationary or drifting data.
Taking this research one step further, Polikar also is considering the role of unlabeled data. For example, computers are constantly bombarded with potential malware, so human experts are required to analyze and label the data to train a system so that it can identify serious attacks from legitimate network traffic. “So you need a human expert to label the data before you can use such data with any of the algorithms, and human expertise is expensive,” Polikar said.
Polikar and his students have applied for funding from the National Science Foundation for research developing such algorithms that can learn from, track and adapt to changing data without a human expert providing examples of labeled data. “We believe we are one of the first — if not the first — to attack this particular problem,” Polikar said.
To disseminate research on learning in nonstationary and evolving environments, Prof. Cesare Alippi from Milan, Italy, and Polikar (associate editor of the journal) are producing a special issue of the IEEE Transactions on Neural Networks and Learning Systems to be published later this year. The Rowan researcher also organized and co-chaired the 2013 IEEE Symposium on Computational Intelligence in Dynamic and Uncertain Environments in Singapore in April, which provided a forum to researchers from around the world working on similar areas.
If successful in this research, Polikar said, this application could help detect malware and spam but also could have potential in financial fraud detection and predicting energy demand and pricing. “Energy demand — and as a result energy prices — changes constantly at all times, but we only know the price or the demand after the fact. The ability to predict short-, medium- and long-term demand under different scenarios would be extremely beneficial to the energy industry.”
This type of algorithm also could factor in the effects of climate change on meteorologic predictions. “Today we have some understanding of under what conditions a twister will form, but is it possible that if the climate conditions change, the conditions under which these meteorological phenomena will occur may also change?” he said. “We don’t know that. We can’t know that until after the fact. We would have to have labeled data indicating that under these conditions a twister did or did not occur. However, tracking long-term unlabeled and unstructured data, it may be possible to predict consequences of changes in weather patterns before the consequences are realized.”
As Polikar proceeds with this investigation, he continues to pursue additional types of research.
His earlier research focused on early detection of Alzheimer’s disease (AD), examining the value of electroencephalogram (EEG) data in early diagnosis in a collaborative study with the University of Pennsylvania and Drexel University in Philadelphia.
Currently, he and his undergraduate and graduate engineering clinic students are examining EEG, magnetic resonance imaging and positron emission tomography data from the University of Pennsylvania. This study recruited patients with normal cognitive function, AD and mild cognitive impairment (MCI), where a subset of the latter group typically develops AD.
Polikar is currently studying whether a suitable fusion of these three types of data – EEG, MRI and PET – can increase the accuracy of early detection. “In addition, we’re also interested in determining whether a person diagnosed with MCI will develop AD,” he said.
# # #