Optimized Privacy-Preserving CNN Inference With Fully Homomorphic Encryption¶
Authors: Dongwoo Kim; Cyril Guyot
Year: 2023
Venue: IEEE Transactions on Information Forensics and Security
type: technique collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
Proposes a coefficient-based homomorphic convolution for CKKS whose cost is independent of the kernel size, giving 12-46x faster convolutions for kernel widths 3-7, and integrates it with CKKS bootstrapping (CoeffToSlot, EvalSine, SlotToCoeff) and high-precision composite-polynomial ReLU. On 20-layer CNN classifiers this yields at least 18.9% lower inference time on CIFAR-10/100 (48.1% on part of an ImageNet network) and, for convolution-heavy CNNs, at least 5x faster FHE inference than prior work with equal or better accuracy.
Abstract¶
Inference of machine learning models with data privacy guarantees has been widely studied as privacy concerns are getting growing attention from the community. Among others, secure inference based on Fully Homomorphic Encryption (FHE) has proven its utility by providing stringent data privacy at sometimes affordable cost. Still, previous work was restricted to shallow and narrow neural networks and simple tasks due to the high computational cost incurred from FHE. In this paper, we propose a more efficient way of evaluating convolutions with FHE, where the cost remains constant regardless of the kernel size, resulting in 12–46× timing improvement on various kernel sizes. Combining our methods with FHE bootstrapping, we achieve at least 18.9% (and 48.1%) timing reduction in homomorphic evaluation of 20-layer CNN classifiers (and a part of it) on CIFAR10/100 (and ImageNet, respectively) datasets. Furthermore, in consideration of our methods being effective for evaluating CNNs with intensive convolutional operations and exploring such CNNs, we achieve at least 5× faster inference on CIFAR10/100 with FHE than the prior works having the same or less accuracy. Index Terms— Privacy-preserving machine learning, fully homomorphic encryption, convolutional neural network.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | kernel-size-independent (constant-time) homomorphic convolution | llm | 0.85 |
| contribution | integration of convolution with CKKS bootstrapping and ReLU evaluation | llm | 0.85 |
| contribution | efficient extraction of valid values for strided convolutions | llm | 0.85 |
| contribution | high-precision polynomial ReLU approximation for deep CNNs | llm | 0.85 |
| contribution | exploration of convolution-intensive CNNs suited to FHE | llm | 0.85 |
| schemes | CKKS | llm | 0.85 |
| libraries | Lattigo | llm | 0.85 |
| models | 20-layer CNN | llm | 0.85 |
| datasets | CIFAR-10 | llm | 0.85 |
| datasets | CIFAR-100 | llm | 0.85 |
| datasets | ImageNet | llm | 0.85 |
| tasks | image classification | llm | 0.85 |
| tasks | private CNN inference | llm | 0.85 |
| security level | 128-bit | llm | 0.85 |
| threat model | semi-honest service provider; client encrypts input (model may also be encrypted) | llm | 0.85 |
| interaction | non-interactive FHE | llm | 0.85 |
| code url | https://github.com/ldsec/lattigo (URL found in the full text) | regex | 0.50 |
| limitations | bootstrapping is required at each layer for high-precision deep CNNs | llm | 0.85 |
| limitations | computing party must know the network structure (number of layers and batches) | llm | 0.85 |
| limitations | scaling activations into [-1, 1] sacrifices some precision | llm | 0.85 |
| llm relevance | medium | llm | 0.85 |
| key result | convolution speed-up: 12-46x (kernel widths 3-7 vs vector-encoding baseline) | llm | 0.85 |
| key result | convolutional layer time reduction: 13-83% (depending on kernel and batch size, with bootstrapping) | llm | 0.85 |
| key result | CNN inference time reduction: at least 18.9% (20-layer CNN on CIFAR-10/100) | llm | 0.85 |
| key result | partial network time reduction: 48.1% (part of a 20-layer CNN on ImageNet) | llm | 0.85 |
| key result | inference speed-up vs prior work: at least 5x (CIFAR-10/100 at same or better accuracy) | llm | 0.85 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 19 [p.1], TFHE 5 [p.2], BGV 3 [p.1], Paillier 2 [p.13], Leveled / somewhat HE 2 [p.2]
- library: EVA / CHET 4 [p.2], Lattigo 2 [p.7], EzPC / CrypTFlow 1 [p.13]
- model: CNN 87 [p.1]
- dataset: CIFAR-10 19 [p.1], ImageNet 11 [p.1], CIFAR-100 3 [p.8], MNIST 2 [p.2]
- hardware: GPU 2 [p.9], AMD EPYC 1 [p.7]
- technique: Convolution 135 [p.1], Bootstrapping 42 [p.1], ReLU 22 [p.2], Rotations 21 [p.1], Compiler 5 [p.2], Polynomial approximation 2 [p.10], RNS 2 [p.6]
- assumption: LWE 3 [p.9]
- task: Language modeling / generation 4 [p.3]
- topic: Homomorphic encryption 92 [p.1], Private / secure inference 11 [p.1], Privacy-preserving machine learning 10 [p.1], Benchmark 2 [p.6]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/inference/kim-2023_optimized-privacy-preserving-cnn-inference-with-fhe.pdf |
ppml/inference | 13 | 5.3 MB | e1ec8385b624 |
was Kim and Guyot - 2023 - Optimized Privacy-Preserving CNN Inference With Fully Homomorphic Encryption.pdf |
Database key: optimized-privacy-preserving-cnn-inference-with-fully-homomorphic · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set optimized-privacy-preserving-cnn-inference-with-fully-homomorphic COLUMN VALUE