Master Degree Project Presentation: CKKS-Based Privacy-Preserving Learning

Abstract: This thesis explores privacy-preserving learning (PPML) via the state-of-the-art homomorphic encryption scheme CKKS. We begin by carefully formalizing the cryptographic scheme as well as the underlying computationally hard problem (Ring Learning with Errors) which offers its semantic security. Subsequently, we focus on approximating various activation functions with low-degree polynomials, our final goal being the construction of models which allow for encrypted training and inference. These approximations are later integrated into encrypted logistic regression and encrypted approximate-SVM via shallow neural networks models. Our results show that CKKS-based privacy learning may achieve practical accuracy without compromising the security of the data, albeit with tradeoffs in time and memory efficiency.