Research map / Federated learning and privacy / Privacy attacks and secure learning
DeepReDuce: ReLU Reduction for Fast 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 |
|---|---|---|---|
| Apply DeepReDuce ReLU optimization to ResNet18 | ResNet18 · CIFAR-100 | top-1 accuracy at matched ReLU budget: 3.5 | Helped |
| Apply DeepReDuce ReLU optimization to ResNet18 | ResNet18 · CIFAR-100 | ReLU count at matched accuracy: 3.5× | Helped |
| Use DeepReDuce instead of CryptoNAS at comparable accuracy | ResNet18 · CIFAR-100 | ReLU count: 100K -> 28.6K | Helped |
| Cull S1 and thin the remaining stages (ablation) | ResNet18 · CIFAR-100 | top-1 accuracy: 73.14 -> 76.22 | Helped |
| Use alternate-layer ReLU thinning rather than channel scaling (ablation) | ResNet18 · CIFAR-100 | top-1 accuracy: 74.72 vs 73.78 | Helped |
| Use fmap scaling rather than channel scaling for the same ReLU reduction (ablation) | ResNet18 · CIFAR-100 | top-1 accuracy with KD: 70.19 -> 72.72 | Helped |
| Add knowledge distillation (ablation) | ReLU-only S3 ResNet18 · CIFAR-100 | top-1 accuracy: 67.41 -> 69.92 | Helped |
| Add knowledge distillation (ablation) | ReLU-only S1 ResNet18 · CIFAR-100 | top-1 accuracy: 61.93 -> 59.85 | Hurt |
| Keep ReLUs in S3 rather than only in Conv1 (ablation) | ResNet18 · CIFAR-100 | top-1 accuracy: 46.22 -> 69.92 | Helped |
| Replace the full network with a shallower ResNet9 (ablation) | ResNet18 · CIFAR-100 | top-1 accuracy: 68.5 vs 66.2 | Hurt |
| Cull S1 and thin S2, S3, and S4 (ablation) | ResNet18 · TinyImageNet | top-1 accuracy: 61.28 -> 62.26 | Helped |
| Use DeepReDuce rather than channel pruning | ResNet56 · CIFAR-100 | top-1 accuracy: 70.83 -> 71.68 | Helped |
Closest papers by meaning
- ReLUPruner: Rethinking ReLU Importance with Taylor Expansion for Efficient Private Inference (AAAI 2026 · similarity 0.92)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.86)
- Disparate Impact on Group Accuracy of Linearization for Private Inference (ICML 2024 · similarity 0.86)
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.85)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.84)
- PrivCirNet: Efficient Private Inference via Block Circulant Transformation (NeurIPS 2024 · 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.