Historically, cryptography was created and used to deliver messages between people in such a way that only the intended receiver would be able to read them. The two persons communicating trusted each other, but not the rest of the world - for instance, not the message bearer. But what happens if we assume instead that even the receiver of a message may not be trusted by the sender, while both still have an interest in communicating? This problem is addressed by secure multi-party computation, that studies communication and shared computation between mutually distrusting parties. In a world where most of the communication happens over the Internet, and where an ever-increasing amount of computation is outsourced, for example to the cloud, multiparty computation is a subject of crucial importance. However, unconditionally secure multi-party protocols have never found practical application: the lack of realistic noisy channel models, that are required to achieve security against computationally unbounded adversaries, prevents implementation over real-world, standard communication protocols.
In this thesis, we study the properties and characteristics of noisy channels, and we introduce new models based on real-world error sources that have never been studied in this context, such as delays. On the contrary, previous works mostly focused on theoretical noisy channels that have little in common with the reality of network communication. We show that the models we propose are suitable for multi-party computation by building on each of them a protocol for oblivious transfer, a fundamental cryptographic primitive that implies any secure computation. Thanks to the realism of our noisy channels, we are able to provide the first working implementation of unconditionally secure oblivious transfer over both wired and wireless networks and different Internet protocols, which we test thoroughly in an extensive set of experiments. Our constructions are not only more realistic, but also more flexible, as they allow larger ranges of error probabilities. This will eventually open the way for the widespread use of unconditionally secure multi-party computation.