Research map / Architectures and efficiency / CNN design
ENSEI: Efficient Secure Inference via Frequency-Domain Homomorphic Convolution for Privacy-Preserving Visual Recognition
Song Bian, Tianchen Wang, Masayuki Hiromoto, Yiyu Shi, Takashi Sato
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use frequency-domain secure convolution instead of Gazelle's time-domain convolution | ENSEI-NTT · Convolution (32 × 32 × 32 input, 3 × 3 × 32 filter) | online time: 195 ms -> 16.9 ms | Improved |
| Use frequency-domain secure convolution instead of Gazelle's time-domain convolution | ENSEI-NTT · Convolution (32 × 32 × 32 input, 3 × 3 × 32 filter) | setup time: 704 ms -> 21.4 ms | Improved |
| Use frequency-domain secure convolution instead of Gazelle's time-domain convolution | ENSEI-NTT · Convolution (28 × 28 × 1 input, 5 × 5 × 5 filter) | online time: 9.20 ms -> 1.8 ms | Improved |
| Use ENSEI with BC3 architecture instead of prior-art secure inference | ENSEI · CIFAR-10, BC3 in [30] | inference time: 35.8 s -> 14.7 s | Improved |
| Use ENSEI with BC4 architecture instead of prior-art secure inference | ENSEI · CIFAR-10, BC4 in [30] | inference time: 123.9 s -> 30.91 s | Improved |
| Replace the second and fifth ReLU layers with square activation (ablation) | ENSEI on the Fig. 13 in [25] architecture · CIFAR-10 | inference time: 7.72 s -> 1.38 s | Improved |
| Replace a small portion of ReLU layers with square activation (ablation) | High-Square · CIFAR-10, same architecture as [23,25] | accuracy: 82% -> 82% | No clear effect |
| Replace all ReLU layers with square activation (ablation) | The square-activation variant · CIFAR-10 | accuracy: 82% -> 10% | Hurt |
| Use binary-weight precision rather than the reported original accuracy setting (ablation) | ENSEI, same architecture as [23,25] · CIFAR-10 | accuracy: 82% -> 81% | Hurt |
| Use high precision (11-bit features and 7-bit weights) rather than full-bit precision (ablation) | ENSEI, same architecture as [23,25] · CIFAR-10 | accuracy: 83% -> 82% | No clear effect |
| Use medium precision (8-bit input and 4-bit filter) rather than full-bit precision (ablation) | ENSEI, same architecture as [23,25] · CIFAR-10 | accuracy: 83% -> 79% | Hurt |
Closest papers by meaning
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.89)
- Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption (ICML 2024 · similarity 0.86)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.86)
- CryptoFace: End-to-End Encrypted Face Recognition (CVPR 2025 · similarity 0.85)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.85)
- PrivCirNet: Efficient Private Inference via Block Circulant Transformation (NeurIPS 2024 · similarity 0.83)
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.
