Research map / Federated learning and privacy / Privacy attacks and secure learning
ULD-Net: Enabling Ultra-Low-Degree Fully Polynomial Networks for Homomorphically Encrypted Inference
Xi Xie, Ran Ran, Jiahui Zhao, Bin Lei, Zhijie Jerry Shi, Wujie Wen, Caiwen Ding
Read the paperCheck 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 ULD-Net quadratic activations with PolyNorm instead of cascaded polynomials (SMART-PAF) | fully polynomial ResNet-18 · ImageNet | top-1 accuracy: 69.40% -> 69.79% | Improved |
| use degree-2 ULD-Net instead of SMART-PAF degree 81 for HE model latency | fully polynomial ResNet-18 · ImageNet | model HE latency (s): 114277 -> 41408 | Improved |
| use degree-2 ULD-Net instead of SMART-PAF degree 81 for activation latency | fully polynomial ResNet-18 · ImageNet | activation latency (s): 8311 -> 1024 | Improved |
| use fully polynomial ULD-Net instead of partial replacement AutoReP | ResNet-18 · CIFAR-100 | top-1 accuracy: 75.48% -> 78.81% | Improved |
| use fully polynomial ULD-Net instead of SNL | ResNet-18 · CIFAR-100 | model HE latency (s): 2052 -> 647 | Improved |
| use ULD-Net instead of NEXUS (low-degree RoPE, PolyNorm, PolyAct) | ViT-Small · CIFAR-10 | non-polynomial operator latency (s): 7995 -> 390 | Improved |
| use ULD-Net instead of NEXUS | ViT-Small · Tiny-ImageNet | top-1 accuracy: 60.52% -> 61.40% | Improved |
| use ULD-Net instead of NEXUS | ViT-Small · CIFAR-10 | top-1 accuracy: 91.39% -> 91.48% | Improved |
| increase PolyAct degree from 2 to 3 (ablation) | ULD-Net VanillaNet-6 · ImageNet | top-1 accuracy: 74.25% -> 76.03% | Improved |
| increase PolyAct degree from 2 to 3 (ablation) | ULD-Net VanillaNet-7 · ImageNet | top-1 accuracy: 74.91% -> 76.40% | Improved |
| increase PolyAct degree from 2 to 3 (ablation) | ULD-Net VanillaNet-7 · ImageNet | model HE latency (s): 5204 -> 5614 | Hurt |
| choose normalization axis per sample (all axes except batch) rather than over the batch (ablation) | fully polynomial models · stability analysis | numerical stability | Improved |
Closest papers by meaning
- Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption (ICML 2024 · similarity 0.90)
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.89)
- PrivCirNet: Efficient Private Inference via Block Circulant Transformation (NeurIPS 2024 · similarity 0.87)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.86)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.86)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.86)
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.
