Description
This thesis explores the integration of Functional Encryption (FE) into Federated Learning (FL)
to enhance privacy guarantees without compromising efficiency or scalability. It presents a
comparative evaluation of four FE-based FL frameworks: HybridAlpha, CryptoFE, 2DMCFE,
and the newly proposed PIM-MCFE, highlighting their architectural design, cryptographic
assumptions, threat models, and performance characteristics. While HybridAlpha achieves
low communication overhead with support for dynamic participation, it suffers from centralized
trust dependencies and decryption inefficiencies. CryptoFE addresses key management
weaknesses through decentralization but introduces significant communication overhead due
to large ciphertexts. 2DMCFE advances the state of the art by eliminating the trusted third
party, enforcing contextual binding through label-constrained decryption, and preventing
mix-and-match attacks, yet still faces performance limitations in large-scale settings. PIMMCFE
further extends these capabilities by enabling intermediate model aggregation with
enhanced flexibility, modularity, and quantization support, offering a compelling trade-off
between accuracy and privacy. Across these evaluations, the thesis sheds light on the core
limitations of current FE-based FL protocols and outlines future directions for building
scalable, privacy-preserving, and context-aware federated learning systems.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching