Research map / Federated learning and privacy / Privacy attacks and secure learning
Selective Network Linearization for Efficient Private Inference
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 |
|---|---|---|---|
| Use SNL instead of DeepReDuce at low ReLU counts | ResNet18 on CIFAR-100 · CIFAR-100 | test accuracy: 68.68 -> 70.05 | Helped |
| Use SNL instead of DeepReDuce at a high ReLU budget | Wide-ResNet 22-8 on CIFAR-100 · CIFAR-100 | test accuracy: 75.51 -> 76.35 | Helped |
| Use SNL instead of DeepReDuce at a comparable ReLU count | ResNet18 on Tiny-ImageNet · Tiny-ImageNet | test accuracy: 57.51 -> 63.39 | Helped |
| Use channel-wise SNL instead of SAFENet | CIFAR-100 private inference · CIFAR-100 | test accuracy: 67.50% vs. 68.26% | Helped |
| Use channel-wise SNL instead of SAFENet | CIFAR-100 private inference · CIFAR-100 | online latency: 0.628s vs. 7.20s | Helped |
| Zero out feature-map outputs instead of passing pre-activated inputs through linearized units (ablation) | SNL on ResNet18 · CIFAR-100 | test accuracy | Hurt |
| Train SNL from scratch rather than starting from a pretrained network (ablation) | SNL on ResNet18 · CIFAR-100 | test accuracy | Mixed |
| Increase the initial Lasso coefficient (ablation) | SNL on ResNet18 · CIFAR-100 | test accuracy | Hurt |
| Use a higher learning rate (0.1 or 0.01) instead of 0.001 (ablation) | SNL on ResNet18 · CIFAR-100 | test accuracy | Hurt |
| Use learning rate 0.001 instead of 0.0001 (ablation) | SNL simultaneous training · CIFAR-100 | test accuracy: 71.55% -> 75.04% | Helped |
| Increase the sparsity-factor schedule too aggressively (ablation) | SNL on ResNet18 · CIFAR-100 | test accuracy | Hurt |
| Gradually increase the Lasso coefficient rather than using a large coefficient immediately (ablation) | SNL on ResNet18 · CIFAR-100 | test accuracy | Helped |
Closest papers by meaning
- DeepReDuce: ReLU Reduction for Fast Private Inference (ICML 2021 · similarity 0.91)
- CryptoNAS: Private Inference on a ReLU Budget (NeurIPS 2020 · similarity 0.89)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.89)
- Disparate Impact on Group Accuracy of Linearization for Private Inference (ICML 2024 · similarity 0.88)
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.87)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.85)
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.