Dissertation Title: Secure Computation in Decentralized Systems

Abstract: 

Decentralized systems like Bitcoin and Ethereum are real-world examples of secure distributed systems deployed at scale. Over the past decade, these systems and others have proven to provide a trust-minimized solution for computing. They ensure the correct execution of code (correctness), maintain the integrity of stored data, and remain consistently available (availability). Additionally, they allow any user to interact without the risk of censorship.

However, while decentralized systems guarantee security properties like integrity, correctness, and availability, they do not provide privacy. In this regard, they are strictly worse than assuming full trust in a centralized server, since any node in the network must see all data. Furthermore, in many of these open systems (also known as 'permissionless' networks), there are no restrictions on who can operate a node. This means that decentralized systems, and public blockchains in particular, cannot operate on private data, greatly limiting the kinds of use-cases they can support.

This dissertation explores solutions to mitigate the privacy concerns associated with modern decentralized systems, focusing particularly on blockchains. The research employs Secure Multiparty Computation (MPC) techniques to address these issues, demonstrating how MPC, which already shares a similar distributed trust threat model, can enhance privacy in decentralized systems. More specifically, this thesis focuses on the following key areas in decentralized systems:

Access Control Mechanisms and Confidential Smart Contracts: The thesis begins by exploring access control mechanisms on blockchains, and from that builds up to the concept of confidential smart contracts -- arbitrary programs that execute both correctly and privately.

Identity Management and Authentication: Building on access control and confidential smart contracts, we examine identity management and authentication within decentralized networks. We develop a highly efficient Threshold ECDSA protocol that runs in the server-aided MPC model.

Perhaps more importantly, we revisit the server-aided MPC model itself, which sits somewhere between the dishonest and honest-majority MPC paradigms, and show that a confidential smart contract is a real-world realization of the server in this model. We thus theorize that dishonest MPC protocols in general can be practically improved under this model, and argue that because there is a real-world counterpart, this model is realistic.

An Improved Distributed Point Function (DPF) and ORAM: A major theoretical contribution of this work is a novel three-party Distributed Point Function (DPF) construction. This leads to state-of-the-art Oblivious RAM (ORAM) and Distributed ORAM (DORAM) protocols, which are important building blocks in MPC.

Privacy-Preserving Digital Currencies: Using this DPF construction, we revisit the problem of privacy-preserving digital currencies, proposing a solution in the account model. This approach challenges the current consensus that privacy in blockchains requires a UTXO model.

Secure Inference with private retrieval: Lastly, the thesis explores how Large Language Models (LLMs) can perform secure inference while retrieving data from private, distributed databases. This method represents a step towards building secure decentralized AI systems that respect user privacy.

Committee members: 

Prof. Alex 'Sandy' Pentland
Toshiba Professor of Media Arts and Sciences
MIT Media Lab

Prof. Srini Devadas
Edwin Sibley Webster Professor of Electrical Engineering and Computer Science
MIT Computer Science and Artificial Intelligence Laboratory

Dr. Thomas Hardjono
Chief Technology Officer
MIT Connection Science and Engineering