Description
This thesis investigates the integration of homomorphic encryption (HE) in federated learning
environments through systematic benchmarking and analysis. This study evaluates the effects
of HE on system performance, model accuracy, privacy preservation, and resilience against
adversarial threats. Beyond baseline evaluation, it explores optimization strategies, including
batching and quantization, as well as hybrid approaches combining HE with differential
privacy and secure multi-party computation. A literature-based analysis further examines
advanced techniques such as hardware acceleration and multi-key schemes. Findings highlight
HE’s strengths in ensuring strong confidentiality and regulatory compliance, while also
revealing significant computational and communication costs and residual privacy risks. By
assessing trade-offs between privacy and performance, the work identifies conditions under
which HE can be practically deployed in federated learning and provides recommendations
for enhancing its efficiency. The thesis concludes with guidance for future research aimed at
achieving a balanced and practical integration of privacy-preserving techniques.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching