Consistent Adaptive Two-Party Computations

Secure protocols for complicated tasks are usually constructed in two phases. Initially, one designs a protocol that is secure in a semihonest model. Second, zero-knowledge correctness proofs are added to assure correctness and privacy against malicious adversaries. Often the corresponding communication and computational overhead makes this approach intractable in practice. Therefore, we define an intermediate notion of security—consistency. In a consistent protocol, participants always learn if their opponent cheats but cannot complain without violating their own privacy. Thus, a potential victim has to choose between the utility of the correct output and a potential privacy breach. In many contexts, where the long-term reputation of a service provider is more important than the privacy of an individual query, such security notion provides adequate protection with a minimal overhead. Private inference control and adaptive oblivious transfer are the most straightforward examples of such protocols.