Towards the AlexNet Moment for Homomorphic Encryption: HCNN, the First Homomorphic CNN on Encrypted Data With GPUs¶
Authors: Ahmad Al Badawi; Chao Jin; Jie Lin; Chan Fook Mun; Sim Jun Jie; Benjamin Hong Meng Tan; Xiao Nan; Khin Mi Mi Aung; Vijay Ramaseshan Chandrasekhar
Year: 2021
Venue: IEEE Transactions on Emerging Topics in Computing
type: system collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
HCNN evaluates convolutional neural networks on BFV-encrypted images with GPU acceleration for non-interactive Deep-Learning-as-a-Service. Low-precision networks (4-bit weights for a 5-layer MNIST CNN, 8-bit for an 11-layer CIFAR-10 CNN) are trained from scratch with straight-through estimators, plaintext CRT decomposition handles the large plaintext modulus, and the networks are implemented in SEAL and a GPU-accelerated BFV library (AFV). At more than 80-bit security it reaches 99% accuracy on MNIST and 77.55% on CIFAR-10 with per-image latencies of 5.16 s and 304.43 s, while classifying batches of over 8,000 images without extra overhead.
Abstract¶
Deep Learning as a Service (DLaaS) stands as a promising solution for cloud-based inference applications. In this setting, the cloud has a pre-learned model whereas the user has samples on which she wants to run the model. The biggest concern with DLaaS is the user privacy if the input samples are sensitive data. We provide here an efficient privacy-preserving system by employing high-end technologies such as Fully Homomorphic Encryption (FHE), Convolutional Neural Networks (CNNs) and Graphics Processing Units (GPUs). FHE, with its widely-known feature of computing on encrypted data, empowers a wide range of privacy-concerned applications. This comes at high cost as it requires enormous computing power. In this article, we show how to accelerate the performance of running CNNs on encrypted data with GPUs. We evaluated two CNNs to classify homomorphically the MNIST and CIFAR-10 datasets. Our solution achieved sufficient security level ( > 80 bit) and reasonable classification accuracy (99) and (77.55 percent) for MNIST and CIFAR-10, respectively. In terms of latency, we could classify an image in 5.16 seconds and 304.43 seconds for MNIST and CIFAR-10, respectively. Our system can also classify a batch of images ( > 8,000) without extra overhead.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | First homomorphic CNN inference accelerated on GPUs | llm | 0.85 |
| contribution | Low-precision (4-bit / 8-bit) network training from scratch to limit ciphertext precision growth | llm | 0.85 |
| contribution | Plaintext CRT decomposition to handle large plaintext moduli for CIFAR-10 | llm | 0.85 |
| contribution | Scaling-factor optimization and FHE parameter estimation from required precision | llm | 0.85 |
| contribution | Batched classification of more than 8,000 images at no extra cost | llm | 0.85 |
| schemes | BFV | llm | 0.85 |
| libraries | SEAL | llm | 0.85 |
| libraries | AFV (GPU-accelerated BFV) | llm | 0.85 |
| libraries | NTL | llm | 0.85 |
| libraries | Tensorpack | llm | 0.85 |
| models | 5-layer CNN (MNIST) | llm | 0.85 |
| models | 11-layer CNN (CIFAR-10) | llm | 0.85 |
| datasets | MNIST | llm | 0.85 |
| datasets | CIFAR-10 | llm | 0.85 |
| tasks | image classification | llm | 0.85 |
| tasks | private CNN inference | llm | 0.85 |
| hardware | GPU (model not stated in excerpt) | llm | 0.85 |
| security level | >80-bit | llm | 0.85 |
| threat model | client-server DLaaS; non-interactive; cloud learns nothing about input, client learns nothing about model beyond f(x) | llm | 0.85 |
| interaction | non-interactive FHE | llm | 0.85 |
| code url | https://github.com/tensorpack/ (URL found in the full text) | regex | 0.50 |
| limitations | 80-bit security is a legacy security parameter accepted for performance reasons | llm | 0.85 |
| limitations | CIFAR-10 HCNN implemented only with the GPU library because CPU (SEAL) evaluation would take too long | llm | 0.85 |
| llm relevance | medium | llm | 0.85 |
| key result | accuracy: 99% (MNIST, 4-bit weight CNN, encrypted inference) | llm | 0.85 |
| key result | accuracy: 77.55% (CIFAR-10, 8-bit weight CNN, scalar encoding) | llm | 0.85 |
| key result | latency: 5.16 s per image (MNIST on GPU) | llm | 0.85 |
| key result | latency: 304.43 s per image (CIFAR-10 on GPU) | llm | 0.85 |
| key result | batch size: > 8,000 images without extra overhead (SIMD batching) | llm | 0.85 |
| key result | security level: > 80-bit (BFV parameters) | llm | 0.85 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: BFV 19 [p.3], MPC 8 [p.2], Leveled / somewhat HE 5 [p.1], Garbled circuits 4 [p.2]
- library: Microsoft SEAL 24 [p.7]
- model: CNN 34 [p.1], AlexNet 15 [p.1], ResNet 1 [p.2]
- dataset: MNIST 63 [p.1], CIFAR-10 45 [p.1], ImageNet 1 [p.13]
- hardware: GPU 50 [p.1], NVIDIA V100 3 [p.10], Intel Xeon 3 [p.10], CUDA 1 [p.13]
- technique: Convolution 35 [p.1], RNS 11 [p.7], ReLU 7 [p.2], Bootstrapping 4 [p.3], Quantization 4 [p.2], Relinearization 3 [p.4], Sigmoid 2 [p.4], SIMD packing 2 [p.4]
- assumption: Lattice-based 6 [p.3], RLWE 2 [p.3]
- task: Medical / healthcare 2 [p.5]
- topic: Homomorphic encryption 81 [p.1], Private / secure inference 3 [p.3], Benchmark 2 [p.10]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/inference/al-badawi-2021_towards-the-alexnet-moment-for-he-hcnn-homomorphic-cnn-on-gpus.pdf |
ppml/inference | 14 | 1.3 MB | c3f8b28dde00 |
was Badawi et al. - 2021 - Towards the AlexNet Moment for Homomorphic Encryption HCNN, the First Homomorphic CNN on Encrypted.pdf |
Database key: towards-the-alexnet-moment-for-homomorphic-encryption-hcnn-the-first · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set towards-the-alexnet-moment-for-homomorphic-encryption-hcnn-the-first COLUMN VALUE