Towards secure bioinformatics services (short paper)

In this paper we show how privacy of genomic sequences can be protected while they are analyzed using Hidden Markov Models (HMM), which is commonly done in bioinformatics to detect certain non-beneficial patterns in the genome. Besides offering strong privacy guarantees, our solution also allows protecting the intellectual property of the parties involved, which makes the solution viable for implementation of secure bioinformatics services. In particular, we show how two mutually mistrusting parties can obliviously run the forward algorithm in a setup where one party knows a HMM and another party knows a genomic string; while the parties learn whether the model fits the genome, they neither have to disclose the parameterization of the model nor the sequence to each other. Despite the huge number of arithmetic operations required to solve the problem, we experimentally show that HMMs with sizes of practical importance can obliviously be evaluated using computational resources typically found in medical laboratories. As a central technical contribution, we give improved protocols for secure and numerically stable computations on non-integer values.