Research map / Federated learning and privacy / Privacy attacks and secure learning
Falcon: Fast Spectral Inference on Encrypted Data
Qian Lou, Wen-jie Lu, Cheng Hong, Lei Jiang
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 |
|---|---|---|---|
| replace second FC layer of LoLa with homomorphic spectral FC (block-circulant, k=8) | LoLa MNIST network · MNIST | latency (s): 2.2 -> 1.2 | Improved |
| same replacement | LoLa MNIST network · MNIST | rotations: 225 -> 120 | Improved |
| same replacement | LoLa MNIST network · MNIST | accuracy (%): 98.95 -> 98.95 | No clear effect |
| replace second convolution layer of LoLa with homomorphic convolution (k=16) | LoLa CIFAR-10 network · CIFAR-10 | latency (s): 730 -> 107 | Improved |
| same replacement | LoLa CIFAR-10 network · CIFAR-10 | rotations: 52975 -> 7674 | Improved |
| same replacement | LoLa CIFAR-10 network · CIFAR-10 | accuracy (%): 76.5 -> 76.5 | No clear effect |
| use block-circulant spectral FC in first FC layer instead of LoLa's | LoLa first FC layer · MNIST | rotations: 151 -> 46 | Improved |
| use tighter encryption modulus (340-bit instead of 440-bit) (ablation) | LoLa · MNIST | latency (s): 2.2 -> 2.1 | Improved |
| use tighter modulus Q (330-bit) for LoLa (ablation) | LoLa · CIFAR-10 | latency (s): 730 -> 730 | No clear effect |
| add homomorphic DFT and inverse DFT to the network (ablation) | Falcon · MNIST | noise / modulus Q (bits): 340 -> 430 | Hurt |
| use block-circulant matrix in FC and convolution (ablation) | Falcon · MNIST | noise bits: ~10-bit | Improved |
| sparse weights and compact encoding versus CryptoNets | nGraph-HE / FCryptoNets · MNIST | latency (s): 205 -> 39.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- CryptoNets — CryptoNets: Applying Neural Networks to Encrypted Data with High Throughput and Accuracy (ICML 2016) · better on 1 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks (ICML 2025 · similarity 0.87)
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.86)
- Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption (ICML 2024 · similarity 0.86)
- PrivCirNet: Efficient Private Inference via Block Circulant Transformation (NeurIPS 2024 · similarity 0.85)
- HEPrune: Fast Private Training of Deep Neural Networks With Encrypted Data Pruning (NeurIPS 2024 · similarity 0.85)
- PrivDNFIS: Privacy-preserving and Efficient Deep Neuro-Fuzzy Inference System (AAAI 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.
