2025 Theses Doctoral

Efficient Distributed Ledger Techniques with Applications to Distributed Machine Learning

Taherpour, Amirhossein

The full potential of large-scale decentralized systems, particularly in the Internet of Things and distributed learning, is currently constrained by the persistent trilemma of security, scalability, and interoperability that limits the adoption of distributed ledger technologies.

This dissertation confronts these challenges by introducing a portfolio of application-driven architectures and consensus mechanisms, beginning with the fortification of the foundational ledger layer. This includes a high-throughput coded blockchain tailored for the resource constraints of IoT devices and a hybrid architecture that merges sharding with a directed acyclic graph to deliver the fast, accurate consensus required by both real-time data streams and iterative machine learning updates. Building on this foundation, the work presents a scalable interoperability framework for the Web3 vision, using a directed acyclic graph as a common communication substrate for multiple blockchains.

The practical applicability of these distributed ledger technologies concepts are then demonstrated through two secure, decentralized federated learning frameworks that directly address the bottlenecks in distributed learning: one integrates a directed acyclic graph sidechain ledger with zero-knowledge proofs for privacy-preserving model validation, and the other synergizes a peer-to-peer gossip protocol with a directed acyclic graph-based control plane using virtual voting to achieve Byzantine fault tolerance.

Collectively, this research delivers a portfolio of validated, high-performance frameworks that are not just theoretical constructs, but application-driven solutions designed to enable the next generation of secure, scalable, and interoperable decentralized systems for IoT and distributed learning.