Research map / Federated learning and privacy / Privacy attacks and secure learning
Scalable Privacy-Preserving Neural Network Training over Z2k via RMFE-Based Packing and Mixed-Circuit Computation
Hengcheng Zhou
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 |
|---|---|---|---|
| add mixed-circuit (bit-wise comparison) truncation instead of arithmetic-only truncation | RMFE-packed Galois ring training · 13-bit truncation, rho=1, LAN online | online time (ms): 67.42 -> 17.83 | Improved |
| use PiTrun instead of PiEdabitsR (rho=1) | truncation protocol · LAN truncation | online time: offline 0.63 -> 0.85 s slower; online 64.40 -> 17.83 ms | Mixed |
| use the ours framework instead of Zhou 2025a (same packing density) | end-to-end training Network B · Network B LAN, max pooling (2,3) | training time (h): 12.19 -> 3.86 | Improved |
| use ours instead of Zhou 2025a on Network A in WAN | end-to-end training · Network A WAN, max pooling (1,3) | WAN time (h): 36.74 -> 34.78 | Improved |
| switch max pooling to average pooling (ablation) | Ours (2,3), Network B · Network B LAN | training time (h): 3.86 -> 1.88 | Improved |
| increase packed secrets rho from 1 to 2 (d=3) (ablation) | Ours, Network B, max pooling · Network B LAN | training time (h): 7.84 -> 3.86 | Improved |
| use (3,5)-RMFE instead of (2,d)-RMFEs for 8<=n<=15 parties on Network A (ablation) | Ours · Network A scalability | training time | Hurt |
| use (3,5)-RMFE instead of (2,d)-RMFEs for n>=16 parties (ablation) | Ours · scalability, Networks A-C | training time | Improved |
| use (3,5)-RMFE instead of (2,d)-RMFEs on CNNs (ablation) | Ours · Networks B and C scalability | communication | Improved |
| raise fixed-point precision f from 13 to 20 (ablation) | ciphertext training · Network A MNIST, 15 epochs | accuracy (%): 52.63 -> 91.78 | Improved |
| raise fixed-point precision from f=13 to f=16 (ablation) | ciphertext training · Network A MNIST, 15 epochs | accuracy (%): 52.63 -> 80.11 | Improved |
Closest papers by meaning
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.87)
- ReBoot: Encrypted Training of Deep Neural Networks with CKKS Bootstrapping (AAAI 2026 · similarity 0.87)
- Revisiting ML Training under Fully Homomorphic Encryption: Convergence Guarantees, Differential Privacy, and Efficient Algorithms (ICML 2026 · similarity 0.86)
- PrivDNFIS: Privacy-preserving and Efficient Deep Neuro-Fuzzy Inference System (AAAI 2025 · similarity 0.85)
- FairProof : Confidential and Certifiable Fairness for Neural Networks (ICML 2024 · similarity 0.84)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.84)
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.
