There is obviously a great deal of good in data analysis - we will soon be able to successfully predict outbreaks, rather than creating computer models that predict the past fine but perform poorly in the nonlinear real world. Even now, large databases like those of the U.S. Census Bureau contain data that, when aggregated and analyzed, can point possible social, economic and health problems and solutions, but the old standby of removing identifying data from files before publishing databases and analytic results doesn't work when there are multiple open public databases available and information data can easily be correlated between databases to pull together bits and pieces of deleted data and recover the identifying information.
One example was a database of movies from a paid service. The database owners wanted to be able to predict what people wanted to view, so they set up a competition to develop the best learning algorithms for real data. They removed identifying information but when viewing patterns from the paid service were compared to viewing patterns on a public, self-reporting movie-viewing database, there was enough information to identify an overwhelming majority of the individuals involved.
Suddenly, anyone could know what movies you watched, including those that people might prefer others did not know they viewed.
A visualization of the spreading of messages on Twitter (retweets network in green) on the followers network (grey). The nodes represent users and their size is proportional to the number of followers that they have. Red indicates users who have written original tweets and yellow indicates users who have retweeted them. Image adapted by A.J. Morales, R.M. Benito et al.-Social Networks
Sofya Raskhodnikova, associate professor of computer science and engineering at Pennsylvania State University, suggests that "differential privacy" is needed to maximize accuracy of analysis while preventing identification of individual records. Differential privacy restricts the types of analyses that can be performed to those for which the presence or absence of one person does not matter much. It guarantees that an analysis performed on two databases that differ in only one record will return nearly the same result.
"One approach for achieving differential privacy is adding a small amount of noise to the actual statistics before publishing them," said Raskhodnikova. "One problem is figuring out how much noise -- it depends on how sensitive the statistic is -- and how to do it so that we can still retain accuracy of results. There are other more sophisticated approaches for achieving differential privacy."
The notion of differential privacy may be especially important to protection of graph data.
Publishing a network of romantic relationships between students in one high school using only gender as an identifier would seem innocuous, for example, but in reality is not. A person with knowledge of the school under study could easily correlate the observed behavior of an outlier - a student with a minority sexual preference, for example - to identify that student's node within the network. Once one node is linked to a particular person, only a small amount of additional information is required to identify the rest of the students included in the network.
As long as the class and students are not identified, researchers once thought that protecting this type of information was not important. However, they now know that using other databases that might be open and available online and various clues from the graph such as the total number of students and the ratio of boys to girls, it might be possible to identify the class and then individual students.
"What the researchers tried to do is strip identifiers -- name, etc., and publish, but this is not enough," said Raskhodnikova. "Other information is available that can be correlated with the network information."
Researchers have found differentially private methods for releasing many graph statistics. Examples include the number of occurrences of specified small patterns in a graph and the degree distribution. The degree distribution of a social network specifies what fraction of people have no friends, one friend, two friends, etc. However, some information, for example, the exact number of connections a specific person has, is inherently too sensitive to be released with differential privacy.
Data that classifies as an outlier also poses a problem. If only one person in a salary database makes over $1 million a year, it becomes fairly easy to identify that person. Some types of information inherently pose a threat to privacy, said Raskhodnikova, and should not be published.
Comments