Secured Anonymous Federated Learning for Privacy-Preserving Pervasive Healthcare

As global reliance on data-driven technologies intensifies, digital privacy concerns have escalated—fuelled by recurring data breaches, large-scale surveillance, and misuse of personal information. Federated learning offers a decentralised alternative that keeps data on local devices, reducing many traditional risks. However, decentralisation alone does not ensure full privacy: gradients, participation timing, and metadata can still reveal sensitive information, leaving clients vulnerable to identification or inference attacks.

To address these gaps, this thesis presents Anonymous Federated Learning (AFL)—a framework designed to protect both data and the identity of its source. AFL uses staged proxy-based routing across key exchange, model download, and gradient upload to break the link between clients and their updates. Implemented with lightweight cryptographic primitives and browser-level interoperability, AFL maintains model accuracy while remaining suitable for heterogeneous and resource-limited devices.

AFL is assessed through extensive experiments covering architectural design, performance benchmarking, and security evaluation. These studies analyse trade-offs in accuracy, computational and communication overhead, and scalability, and benchmark AFL against Differential Privacy, Secure Aggregation, and Homomorphic Encryption. Further investigations explore browser-side optimisation, compression, GPU acceleration, and hybrid privacy configurations.

The findings show that AFL provides competitive accuracy and efficiency while offering stronger anonymity than existing techniques. It performs particularly well in settings with high client diversity or constrained resources, where heavier cryptographic approaches become impractical. Combining AFL with complementary mechanisms such as Differential Privacy or compression further improves robustness and scalability.

In summary, this work delivers a practical and experimentally validated approach to privacy-preserving distributed learning. By shifting the focus from protecting what is shared to concealing who shares it, Anonymous Federated Learning offers a new foundation for secure, anonymous, and equitable participation in federated AI systems.