BERKELEY, Calif. – A handful of Berkeley researchers are part of a team that won a $10 million grant to try to head of hackers at the social networking pass. The National Science Foundation awarded the grant for a five-year project exploring how to anticipate and prevent attacks on social networking Web sites.
The work is just one example of dozens of projects at the International Computer Science Institute (ICSI), a group allied with but not a part of the nearby University of California at Berkeley. As many as 150 researchers work at ICSI on projects that range from coming up with new algorithms for voice recognition and video search to writing open source code for networking gear.
"More and more we are looking at the deep meaning of words, data and what’s happening in the network," said Roberto Pieraccini, ICSI’s director and a former voice recognition researcher. "Twenty years ago we were happy to ID things, today we are moving to a different level--identifying the meaning of things," Pieraccini said.
In the new security program, researchers will try to anticipate and block possible social networking attacks, some of which are already emerging. For example, some hackers have found a business accumulating and selling large numbers of Twitter followers.
In network security, researchers have evolved from seeking technical to economic solutions.
ICSI director Roberto Pieraccini oversees research identifying "the meaning of things."
“Our strong suspicion is today’s spam-email economy, for example, won’t go away but won’t prosper further, so it may decline,” said one researcher involved in the work. “The people involved will probably try to find some other niche, and our bet is heavily on social networking as their platform, and hopefully we will be more ahead of the game this time—at least we have a head start,” the researcher said.
Another researcher, Nelson Morgan, is one of several at ICSI working on new approaches to voice recognition. Today voice recognition products such as Apple’s Siri are based on the same Hidden Markov models that have been used for 15 years, but “there are fundamental flaws as people currently use them,” said Morgan.
“You have to make certain mathematical assumptions that people know are wrong, but you cover that up by using huge amounts of statistical data and limiting the domain,” Morgan said.
The difficulty with the current approach is it cannot perform many functions of human listeners such as blocking out ambient noise or accounting for strong accents. Now researchers are trying to find fresh approaches based on recorded signals of brain patterns and using the parallel computing capabilities of multicore processors.
The new approaches hold promise for what today is “a brittle technology,” said Morgan. “We have some new algorithms already, but most of them are still unknown,” he said.