DOReN: Toward Efficient Deep Convolutional Neural Networks with Fully Homomorphic Encryption¶
Authors: Souhail Meftah; Benjamin Hong Meng Tan; Chan Fook Mun; Khin Mi Mi Aung; Bharadwaj Veeravalli; Vijay Chandrasekhar
Year: 2021
Venue: IEEE Transactions on Information Forensics and Security
type: technique collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
DOReN is a low-depth batched homomorphic neuron, built on the BGV scheme in HElib, that evaluates many quantized ReLU-activated neurons on encrypted data exactly (no polynomial approximation). A depth-optimized accumulator reduces the circuit depth from O(log m * log n) to O(log m + log n) for n-bit integers, with bootstrapping applied dynamically from HElib's capacity estimate and Benes-network permutations for inter-layer transforms. The amortized neuron cost is about 1.26 s for 300 inputs and under 0.13 s for 10 inputs at 80-bit security, roughly 20x better than Lou and Jiang (NeurIPS 2019).
Abstract¶
Fully homomorphic encryption (FHE) is a powerful cryptographic primitive to secure outsourced computations against an untrusted third-party provider. With the growing demand for AI and the usefulness of machine learning as a service (MLaaS), the need for secure training and inference of artificial neural networks is rising. However, the computational complexity of existing FHE schemes has been a strong deterrent to this. Prior works suffered from accuracy degradation, lack of scalability, and ciphertext expansion issues. In this paper, we take the first step towards the problem of space-efficiency in evaluating deep neural networks through designing DOReN: a low depth, batched neuron that can simultaneously evaluate multiple quantized ReLU-activated neurons on encrypted data without approximations. Our circuit design reduced the complexity of the accumulator circuit depth from O(log m · log n) to O(log m + log n) for n bit integers. The experimental results show that the amortized processing time of our homomorphic neuron is approximately 1.26 seconds for 300 inputs and less than 0.13 seconds for 10 inputs at 80 bit security, which is a 20 fold improvement upon Lou and Jiang, NeurIPS 2019.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | batched homomorphic neuron that evaluates multiple quantized ReLU neurons without approximation | llm | 0.80 |
| contribution | depth-optimized accumulator reducing depth from O(log m * log n) to O(log m + log n) | llm | 0.80 |
| contribution | dynamic bootstrapping driven by HElib ciphertext capacity | llm | 0.80 |
| contribution | inter-layer ciphertext transforms via Benes network permutations | llm | 0.80 |
| schemes | BGV | llm | 0.80 |
| libraries | HElib | llm | 0.80 |
| models | CNN | llm | 0.80 |
| datasets | CIFAR-10 | llm | 0.80 |
| tasks | image classification | llm | 0.80 |
| tasks | private DNN inference with FHE | llm | 0.80 |
| hardware | multi-core CPU platform with 104 threads (model not specified) | llm | 0.80 |
| security level | 80-bit (parameters originally 128-bit, downgraded due to newer attack models) | llm | 0.80 |
| threat model | untrusted third-party server (outsourced computation) | llm | 0.80 |
| interaction | non-interactive FHE | llm | 0.80 |
| code url | https://github.com/tfhe/tfhe (URL found in the full text) | regex | 0.50 |
| limitations | accumulation dominates runtime and the accumulator assignment is not multi-threaded | llm | 0.80 |
| limitations | SIMD benefit tapers beyond 80 inputs when computation exceeds available threads | llm | 0.80 |
| limitations | security parameters downgraded from 128-bit to 80-bit under new attack models | llm | 0.80 |
| llm relevance | medium | llm | 0.80 |
| key result | amortized neuron processing time: ~1.26 s (300 inputs, 80-bit security) | llm | 0.80 |
| key result | amortized neuron processing time: < 0.13 s (10 inputs, 80-bit security) | llm | 0.80 |
| key result | improvement over Lou and Jiang (NeurIPS 2019): 20x (homomorphic neuron evaluation) | llm | 0.80 |
| key result | bootstrapping after one Benes permutation: 9-11 s (amortized ~1 s for size-1024 permutations run in parallel) | llm | 0.80 |
| key result | share of neuron time spent in accumulation: 84-97.5% (all input sizes) | llm | 0.80 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: TFHE 23 [p.2], Leveled / somewhat HE 14 [p.2], BGV 10 [p.2], BFV 6 [p.3]
- library: HElib 6 [p.9], Microsoft SEAL 1 [p.13]
- model: CNN 19 [p.2], AlexNet 1 [p.12], VGG 1 [p.11]
- dataset: CIFAR-10 3 [p.11], ImageNet 3 [p.2], MNIST 2 [p.2]
- hardware: Intel Xeon 1 [p.9], GPU 1 [p.12]
- technique: Bootstrapping 51 [p.2], ReLU 37 [p.1], Convolution 31 [p.1], Quantization 20 [p.1], SIMD packing 12 [p.2], Rotations 4 [p.4], Bitslicing 4 [p.5], Lookup table 3 [p.3]
- assumption: RLWE 2 [p.9], LWE 2 [p.13], Lattice-based 2 [p.13], IND-CPA 2 [p.9]
- task: Medical / healthcare 6 [p.1], Finance 2 [p.1]
- topic: Homomorphic encryption 41 [p.1]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/inference/meftah-2021_doren-efficient-deep-cnns-with-fhe.pdf |
ppml/inference | 13 | 1.4 MB | f2ffc350de9b |
was Meftah et al. - 2021 - DOReN Toward Efficient Deep Convolutional Neural Networks with Fully Homomorphic Encryption.pdf |
Database key: doren-toward-efficient-deep-convolutional-neural-networks-with-fully · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set doren-toward-efficient-deep-convolutional-neural-networks-with-fully COLUMN VALUE