Reverse Bayesian poisoning: How to use spam filters to manipulate online elections

E-voting literature has long recognised the threat of denial-of-service attacks: as attacks that (partially) disrupt the services needed to run the voting system. Such attacks violate availability. Thankfully, they are typically easily detected. We identify and investigate a denial-of-service attack on a voter's spam filters, which is not so easily detected: reverse Bayesian poisoning, an attack that lets the attacker silently suppress mails from the voting system. Reverse Bayesian poisoning can disenfranchise voters in voting systems which rely on emails for essential communication (such as voter invitation or credential distribution). The attacker stealthily trains the voter's spam filter by sending spam mails crafted to include keywords from genuine mails from the voting system. To test the potential effect of reverse Bayesian poisoning, we took keywords from the Helios voting system's email templates and poisoned the Bogofilter spam filter using these keywords. Then we tested how genuine Helios mails are classified. Our experiments show that reverse Bayesian poisoning can easily suppress genuine emails from the Helios voting system.