By Joseph Hall on 23 Jun 2011.
Researchers and close ACCURATE confidants at Princeton’s Center for Information Technology Policy, Will Clarkson, Joe Calandrino and Ed Felten, have just released a neat new result (“New Research Result: Bubble Forms Not So Anonymous”).
The central idea in this result is that these researchers have examined how people fill in bubble forms, like optical scan ballots in voting, to see if there is enough structure in these bubble patterns to uniquely identify the individual filling out the form. They apply some serious machine-learning mojo and can correctly identify the individual about 50% of the time, a much greater identification rate than the 3% rate for making completely random guesses. And the correct answer is one of the top three results 75% of the time.
This has both good and bad consequences for elections. Bad in that anyone with form-filling data such as an employer or an exit pollster, likely has enough identifying information to identify a person’s ballot based solely on a scanned image of that ballot, the likes of which advocates (such as the Humboldt Election Transparency Project) have been releasing for a few years now. Good in that this might help to identify when a different person filled out a ballot (vote buying) or, more importantly, if many ballots were filled out by the same person (ballot box stuffing).
The Princeton team has had this paper accepted to USENIX Security in August and they’ve been playing around with mitigations for voting, such as the inked markers used in Los Angeles for the InkaVote system (where a cheap inked dauber can apply a uniform size and amount of ink to a target).
Full disclosure: the author of this post, Joseph Lorenzo Hall, was a visiting postdoc at CITP for the past three years and consulted closely with the CITP team on this work.