Research map / Architectures and efficiency / CNN design
Low-Complexity Deep Convolutional Neural Networks on Fully Homomorphic Encryption Using Multiplexed Parallel Convolutions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| multiplexed parallel convolution (MULTPARCONV) instead of MULTCONV (ablation) | convolution in ResNet-20 · ResNet-20 inference | number of rotations: 4,360 -> 1,657 (62% fewer) | Helped |
| imaginary-removing bootstrapping before approximate ReLU | ResNet-110 inference · simulation, 50 test images | catastrophic divergence occurrence: divergence in 12 of 50 images (24%) removed | Helped |
| full proposed pipeline (multiplexed packing, parallel convolution, imaginary-removing bootstrapping, level optimization, lighter parameters) | ResNet-20 on CIFAR-10 · CIFAR-10 latency | runtime (single thread vs 64-thread prior work): 10,602s -> 2,271s (4.67x) | Helped |
| full proposed pipeline on ResNet-20 | ResNet-20 CIFAR-10 · amortized runtime per image | amortized runtime (s): 10,602s -> 79s (134x) | Helped |
| proposed pipeline with multiplexed packing and optimized parameters | ResNet-20 bootstrapping KSO count · CIFAR-10 ResNet-20 | number of KSOs: 384,160 -> 3,306 | Helped |
| level optimization by integrating batch normalization into convolution (MULTPARCONVBN) | convolution + batch normalization · level consumption | levels saved: saves three levels | Helped |
| proposed architecture on ResNet-20/32/44/56/110 accuracy vs backbone | CIFAR-10 and CIFAR-100 ResNets on RNS-CKKS · CIFAR-10 accuracy | accuracy vs backbone: within about 0.2 to 0.6 points of backbone in most models (ResNet-110 is 0.55 higher than backbone) | No clear effect |
| lighter and tighter parameters (Hamming weight 192, 51-bit primes, 1,501-bit modulus) | RNS-CKKS parameter set · security level | security: standard 128-bit security achieved (previous work below 128 bits) | Helped |
Closest papers by meaning
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.86)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.85)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.84)
- Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption (ICML 2024 · similarity 0.83)
- PrivCirNet: Efficient Private Inference via Block Circulant Transformation (NeurIPS 2024 · similarity 0.83)
- HEPrune: Fast Private Training of Deep Neural Networks With Encrypted Data Pruning (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.