IBM unveils Web privacy work

Network World - Researchers at IBM's Privacy Institute are working on software that automatically scrambles Web visitors' personal information so consumers won't feel compelled to lie to protect their privacy.
It's no secret that online visitors often provide false personal data to avoid any repercussions should the data be misused or shared with multiple sources. For merchants, that means the customer data they painstakingly track with customer relationship management software -- and often rely on when making product development and marketing decisions -- can be flawed from the start.
To help solve this problem, researchers Rakesh Agrawal and Ramakrishnan Srikant are developing what IBM calls "privacy-preserving data mining." The duo's research, which IBM announced yesterday, relies on the notion that a Web visitor's personal data can be protected if the information is scrambled, or randomized, before it gets to the merchant. Once the data is transferred to the merchant's systems, the IBM software applies algorithms to compensate for the data scrambling. With this technology, a retailer could still generate accurate data models and extract useful demographic information, but without ever seeing personal consumer data, IBM said.
"Our research institutionalizes the notion of fibbing on the Internet, and does so to preserve the overall reality behind the data," Agrawal said.
When a Web user submits a piece of personal data, such as age or salary, the IBM software immediately scrambles that number by adding to or subtracting from it a random value. This randomization step is performed independently for every user, IBM said. This means a 30-year-old's age may be changed to 42, while a 34-year-old's age may become 28.
The merchant determines the range of the randomization -- plus or minus 1 to 12 years, for example -- which then remains constant. Once all the scrambled data is collected for a large number of users, IBM's data mining software determines how the true data might have looked and uses the reconstruction to build a data-mining model, IBM said.
The greater the range of number scrambling that's allowed, the more consumers' private data is obscured. However, as randomization parameters increase, the accuracy of the post-scramble data mining results decreases. According to Agrawal, it's a trade-off. IBM said that in its experiments, after compensating for the data scrambling, it found only a 5% to 10% loss in accuracy, even with 100% randomization allowances.
The research project is under way at IBM's Privacy Institute in Almaden, Calif. Beta trials will begin soon. It's the first project announced by the group, which was formed in November 2001.
Internet privacy is a hot topic, most recently making headlines when U.S. Sen. Ernest "Fritz" Hollings (D-S.C.) introduced a controversial bill designed to safeguard Internet users' privacy, and which opponents suggest will hamper online commerce (see story).

For more coverage of privacy issues, see our privacy page.