Advancing IoT security and privacy: federated learning strategies for smart environments

This doctoral thesis investigates the integration of federated learning (FL) into Internet of Things (IoT) environments, with the aim of enhancing both security and privacy across distributed smart systems. Considering the growing prevalence of data-centric applications within smart homes and other IoT domains, this research addresses the critical challenges posed by data exposure, adversarial threats, and the computational limitations of edge devices. To this end, I propose and develop ConfidenShuffle, a novel federated learning framework that combines encrypted gradient shuffling, differential privacy, and trust-based evaluation mechanisms. This framework is specifically designed to obscure the provenance of client updates, mitigate gradient inversion and poisoning attacks, and ensure the verifiability of model contributions. A key innovation of the approach lies in its integrated trust management scheme, which scores and filters client updates based on behavioural patterns and consistency, thus reducing the influence of unreliable or malicious participants. The framework has been evaluated through simulations using the non-IID MNIST dataset under various adversarial conditions, including the presence of Byzantine and data poisoning attacks. The results demonstrate that ConfidenShuffle maintains high model accuracy while significantly reducing privacy leakage and communication overhead. Moreover, the system exhibits strong resilience to gradient inversion attacks and performs effectively under constrained communication and computation settings typical of smart IoT devices. This research contributes to the field in four significant ways: (1) the design of a privacy-preserving, trust-aware federated learning architecture; (2) the development of defence mechanisms against advanced adversarial threats; (3) the adaptation of FL to non-IID data environments; and (4) the validation of the proposed methods in practical smart home and smart healthcare scenarios. Collectively, these contributions advance the state of secure, scalable, and privacy-conscious learning systems suitable for real-world IoT deployments.