Research map / Architectures and efficiency / RNNs and normalization
PrivCirNet: Efficient Private Inference via Block Circulant Transformation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PrivCirNet, a protocol/network co-optimization for hybrid HE/MPC private DNN inference that uses block circulant weight transformation, a new HE encoding (CirEncode), loss-aware initialization and latency-aware ILP block-size search, plus layer fusion.
Problem
HE-based private DNN inference has very high linear-layer computation cost (HE multiplications and rotations), making latency orders of magnitude higher than plaintext inference; prior HE encodings and pruning are not fully compatible with efficient BSGS rotation schemes.
Approach
Replace dense weight matrices with block circulant matrices so each circulant block becomes a 1D convolution computed by a single HE-Pmult with coefficient encoding; a new nested encoding (CirEncode) combines coefficient and SIMD encodings across blocks to cut HE-Pmult by factor b and HE-Rot by sqrt(b); layer-wise block sizes are chosen by an ILP over second-order loss sensitivity.
Key results
- On Tiny ImageNet, PrivCirNet reduces latency vs Bolt by 1.9x (MobileNetV2), 5.0x (ResNet-18) and 1.3x (ViT) at iso-accuracy.
- Versus SpENCNN, PrivCirNet improves accuracy by 5.2% (MobileNetV2), 4.1% (ResNet-18) and 12% (ViT) on Tiny ImageNet.
- Versus Neujeans and Bolt on GEMM/conv microbenchmarks, PrivCirNet (b8) reduces HE-Rot by about 7x and HE-Pmult by 8.5x on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use block circulant transformation with CirEncode | MobileNetV2 · Tiny ImageNet | latency versus Bolt: 1.9× latency reduction | Helped |
| use block circulant transformation with CirEncode | ResNet-18 · Tiny ImageNet | latency versus Bolt: 5.0× latency reduction | Helped |
| use block circulant transformation with CirEncode | ViT · Tiny ImageNet | latency versus Bolt: 1.3× latency reduction | Helped |
| use block circulant transformation with CirEncode | MobileNetV2 · ImageNet | accuracy versus SpENCNN: 4.2% | Helped |
| use CirEncode with block size 8 | GEMM and convolution microbenchmarks · GEMM and convolution microbenchmarks | HE rotations: 70 -> 10 | Helped |
| use CirEncode with block size 8 | GEMM and convolution microbenchmarks · GEMM and convolution microbenchmarks | HE-Pmult: 2504 -> 155 | Helped |
| replace uniform or Frobenius block-size assignment with loss-aware assignment (ablation) | MobileNetV2 and ViT · CIFAR-10, CIFAR-100, and Tiny ImageNet at 50%, 25%, and 12.5% latency limitations | Top-1 accuracy | Helped |
| remove CirEncode (ablation) | PrivCirNet on MobileNetV2 · Tiny ImageNet | latency | Hurt |
| remove CirEncode (ablation) | PrivCirNet on MobileNetV2 · Tiny ImageNet | accuracy | Hurt |
| add latency-aware block-size assignment (ablation) | PrivCirNet on MobileNetV2 · Tiny ImageNet | accuracy | Helped |
| add fusion methods (ablation) | PrivCirNet on MobileNetV2 · Tiny ImageNet | latency | Helped |
| add fusion methods (ablation) | PrivCirNet on MobileNetV2 · Tiny ImageNet | accuracy | No clear effect |
Surprising findings
- Naively applying existing HE encodings to block circulant matrices gives limited or even increased latency, so encoding co-design is needed.
- Structured pruning is hard to make compatible with BSGS rotation reduction.
Limitations
- Nonlinear (ReLU) layers become the bottleneck after linear-layer speedup; the paper only shows a preliminary ReLU-pruning extension.
- Accuracy degradation grows at higher compression ratios, especially for networks with fewer standard convolution parameters (e.g., MobileNetV2, RegNet).
- Self-attention layers in ViT have no weight matrices and cannot benefit from circulant transformation.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny ImageNet ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny ImageNet ImageNet
Metrics: top-1 accuracy latency (s) HE-Rot count HE-Pmult count
Baselines: Bolt Neujeans Cheetah CrypTFlow2 Falcon SpENCNN
Builds on: Bolt Cheetah Neujeans SpENCNN CirCNN
Closest papers by meaning
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.87)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.87)
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.86)
- Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption (ICML 2024 · similarity 0.85)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.85)
- HEPrune: Fast Private Training of Deep Neural Networks With Encrypted Data Pruning (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.