Research map / Federated learning and privacy / Privacy attacks and secure learning
CryptoNAS: Private Inference on a ReLU Budget
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 |
|---|---|---|---|
| ReLU-shuffle the convolution and skip sequence to reuse activations (ablation) | CNet2 · CIFAR-10 and CIFAR-100 | accuracy: without compromising accuracy | No clear effect |
| Remove ReLUs introduced by skip connections (ablation) | CNet2 with ReLU shuffling · CIFAR-100, 100K ReLU budget | accuracy: 2.2% higher accuracy | Helped |
| Remove ReLUs introduced by skip connections (ablation) | CNet2 with ReLU shuffling · CIFAR-10 and CIFAR-100, 1M ReLU budget | accuracy: 0.64% | Hurt |
| ReLU-shuffle skip connections instead of the ENAS baseline arrangement (ablation) | CNet2 · CIFAR-10 and CIFAR-100 | ReLU count: about 1.9× | Helped |
| Use ReLU-balanced channel scaling rather than FLOP-balanced scaling (ablation) | CNNs under a fixed ReLU budget · CIFAR-100, CNet1, 86K ReLU budget | accuracy: 3.75% | Helped |
| Use ReLU-balanced channel scaling rather than FLOP-balanced scaling (ablation) | CNNs under a fixed ReLU budget · CIFAR-10 and CIFAR-100 | accuracy | Helped |
| Use ReLU-balanced channel scaling rather than FLOP-balanced scaling (ablation) | Wide ResNet · CIFAR-10 | accuracy | Helped |
| Search skip connections with ENAS instead of using the largest all-skip architecture (ablation) | CNet3 · CIFAR-10 | accuracy: 0.24% | Helped |
| Search skip connections with ENAS instead of using the largest all-skip architecture (ablation) | CNet3 · CIFAR-100 | accuracy: 0.48% | Helped |
| Use CryptoNAS rather than ResNet-20 | CIFAR-10 private inference · CIFAR-10 | inference latency: 2.2× faster | Helped |
| Use CryptoNAS rather than DenseNet-12-40 | CIFAR-100 private inference · CIFAR-100 | inference latency: 1.5× faster | Helped |
| Use CryptoNAS rather than Delphi | CIFAR-10 and CIFAR-100 private inference · Smallest 50K ReLU budget | accuracy | Hurt |
Closest papers by meaning
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.85)
- CryptoMoE: Privacy-Preserving and Scalable Mixture of Experts Inference via Balanced Expert Routing (NeurIPS 2025 · similarity 0.85)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.84)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.84)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.84)
- Disparate Impact on Group Accuracy of Linearization for Private Inference (ICML 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.