Enhancing federated learning through distributed ledger technology integration

In today's data-driven world, vast amounts of information power Machine Learning (ML) models for a wide range of applications. However, this data flow raises significant privacy concerns, as individuals are often reluctant to share personal information, especially given increasing regulations on data protection. Federated Learning (FL) offers a solution by training ML models directly on users' devices and sending only model updates to a central server. This distributed approach enables collaboration without sharing personal data, but challenges remain. Centralization may lead to server bottlenecks, reduced resilience, and fairness concerns if updates from certain devices are prioritized. Additionally, the lack of transparency and accountability can erode trust, while security risks, such as data poisoning and model inversion attacks, further complicate FL. Deployment can be costly and time-consuming, and participants may also lack incentives. Regulatory compliance, such as ensuring the right to be forgotten, adds complexity, as removing data from FL models without full retraining is challenging. This dissertation proposes integrating Distributed Ledger Technologies (DLTs) with FL to address these challenges. DLT decentralizes the aggregation process, enhancing security, transparency, and fairness through immutable record-keeping and traceability. Two DLT-based architectures are presented: one blockchain-based and the other using a Directed Acyclic Graph (DAG) for scalability. These approaches utilize Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs) to track contributions and verify participants. Furthermore, a DLT-based FL as a Service (FLaaS) is introduced to simplify deployment, incorporating model validation to mitigate poisoning attacks and token-based incentives to encourage participation. Additionally, this dissertation outlines design guidelines for Federated Unlearning (FU), covering key evaluation metrics, existing techniques, and future research. Finally, a new unlearning algorithm is proposed to address adversarial settings and protect model integrity. These contributions pave the way for more secure, transparent, and resilient FL systems that can meet the needs of next-generation data-driven applications.