Low-latency linear transformations with small key transmission for private neural network on homomorphic encryption
Status. Submitted to Neurocomputing in October 2025 — currently under review (2nd round, minor revision).
Summary. Homomorphic encryption (HE) offers strong security guarantees, but its practical use is limited by high computational latency. In HE convolution, a major source of this cost is the rotation operation and the large rotation keys it requires. This work starts from the observation that multiplexed parallel convolution (MPConv), the state-of-the-art method for HE convolution, incurs a high number of rotation operations and a correspondingly large rotation-key volume.
This work proposes RO-MPConv, an algorithmic redesign of MPConv that reduces convolution latency by up to 81% and rotation-key size by up to 95%, thereby lowering both the computational and the key-transmission overhead of private neural-network inference under HE. Notably, since RO-MPConv is a drop-in replacement for MPConv, simply substituting it into existing state-of-the-art HE models reduces their total inference latency by up to 26%.
My role. I began this work as an undergraduate research intern. The initial conceptualization was suggested by my advisor, Prof. Joon-Woo Lee; building on that, I carried out the experiments, developed and formalized the method (including the diagrams and mathematical formulation), and wrote and revised the manuscript. Homomorphic encryption is difficult to implement, but because this project required engaging deeply with HE algorithms, it helped me get started in the field.