2026-09-19T13:25:52Z https://www.ideals.illinois.edu/oai-pmh

oai:www.ideals.illinois.edu:2142/132598 2026-02-20 col_2142_5131 col_2142_8888 com_2142_5130 com_2142_8887 com_2142_234

application/pdf en text Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2026-02-19 without embargo terms The student, Bill Dai, accepted the attached license on 2025-12-09 at 22:55. The student, Bill Dai, submitted this Thesis for approval on 2025-12-11 at 10:21. This Thesis was approved for publication on 2025-12-11 at 13:37. DSpace SAF Submission Ingestion Package generated from Vireo submission #23117 on 2026-02-19 at 18:30:11 PFHE: partially homomorphic encryption on CNN inference Dai, Bill 2025-12-11 Chen, Deming Homomorphic encryption Machine learning Cybersecurity Privacy-preserved machine learning eng Fully Homomorphic Encryption (FHE) enables secure computation on encrypted data, but its high computational overhead poses significant challenges for practical deep learning inference. In many application scenarios, high-resolution images may only contain a small portion of sensitive information. We noticed that none of the previous works consider this, so in this work, we propose a new framework that accelerates encrypted Convolutional Neural Network (CNN) inference by only encrypting the privacy-sensitive regions of input while processing the remaining parts in plaintext. Our method significantly reduces computational cost without compromising data confidentiality. We developed a new data layout for ciphertexts to utilize the sparse nature of the data. Besides that, we first leverage the capability of FHE to scheme switch between different schemes, such as Cheon-Kim-Kim-Song (CKKS) and Fast Homomorphic Encryption over the Torus (FHEW), achieving the implementation of nonlinear activation like ReLU in CNN with high precision, without the consumption of a significant amount of multiplications. We evaluate our framework under the ImageNet dataset. For the first six convolutional layers, our method achieves at least 4.12× speedup in latency and 28.72× less memory usage compared to the traditional Channel-wise convolution method under various settings. We then present a case study for a hybrid solution, combining our partially FHE encrypted convolution method with Channel-wise convolution, which also shows a theoretical latency reduction of 10.30× and 33.30× less memory consumption. 2025-12 Thesis https://hdl.handle.net/2142/132598 Copyright 2025 Bill Dai Electrical & Computer Eng Electrical & Computer Engr University of Illinois Urbana-Champaign M.S. Thesis