Stealing an Election Paul Kocher (included in the April 2004 issue of Bruce Schneirer's Crypto-Gram)
There are major efforts by computer security professionals to convince government officials that paper audit trails are essential in any computerized voting machine. They have conducted actual examination of software, engaged in letter writing campaigns, testified before government bodies, and collectively, have maintained visibility and public awareness of the issue.
The track record of the computerized voting machines used to date has been abysmal; stories of errors are legion. Here's another way to look at the issue: what are the economics of trying to steal an election?
Let's look at the 2002 election results for the 435 seats in the House of Representatives. In order to gain control of the House, the Democrats would have needed to win 23 more seats. According to actual voting data (pulled off the ABC News website), the Democrats could have won these 23 seats by swinging 163,953 votes from Republican to Democrat, out of the total 65,812,545 cast for both parties. (The total number of votes cast is actually a bit higher; this analysis only uses data for the winning and second-place candidates.)
This means that the Democrats could have gained the majority in the House by switching less than 1/4 of one percent of the total votes -- less than one in 250 votes.
Of course, this analysis is done in hindsight. In practice, more cheating would be required to be reasonably certain of winning. Even so, the Democrats could have won the house by shifting well below 0.5% of the total votes cast across the election.
Let's try another analysis: What is it worth to compromise a voting machine? In contested House races in 2002, candidates typically spent $3M to $4M, although the highest was over $8M. The outcomes of the 20 closest races would have changed by swinging an average of 2,593 votes each. Assuming (conservatively) a candidate would pay $1M to switch 5,000 votes, votes are worth $200 each. The actual value is probably closer to $500, but I figured conservatively here to reflect the additional risk of breaking the law.
If a voting machine collects 250 votes (about 125 for each candidate), rigging the machine to swing all of its votes would be worth $25,000. That's going to be detected, so is unlikely to happen. Swinging 10% of the votes on any given machine would be worth $2500.
This suggests that it is necessary to assume that attacks against individual voting machines are a serious risk.
Computerized voting machines have software, which means we need to figure out what it's worth to compromise a voting machine software design or code, and not just individual machines. Any voting machine type deployed in 25% of precincts would register enough votes that malicious software could swing the balance of power without creating terribly obvious statistical abnormalities.
In 2002, all the Congressional candidates together raised over $500M. As a result, one can conservatively conclude that affecting the balance of power in the House of Representatives is worth at least $100M to the party who would otherwise be losing. So when designing the security behind the software, one must assume an attacker with a $100M budget.
Conclusion: The risks to electronic voting machine software are even greater than first appears.
This essay was written with Paul Kocher.