CrypTen: Secure Multi-Party Computation Meets Machine Learning¶
Authors: Brian Knott; Shobha Venkataraman; Awni Hannun; Shubho Sengupta; Mark Ibrahim; Laurens van der Maaten
Year: 2021
Venue: NeurIPS 2021
type: system collection: catalogue status: auto
Links: arXiv:2109.00984
Summary¶
Auto-extracted summary (unreviewed)
CrypTen is a secure multi-party computation framework that exposes secret-shared tensors, automatic differentiation and modular neural-network layers through a PyTorch-like API, supporting an arbitrary number of semi-honest parties and GPU execution. Benchmarks on text classification (Yelp), Wav2Letter speech recognition and ResNet/ViT image classification show that private inference is practical; two parties can run Wav2Letter faster than real time, although CrypTen remains 2.5-3 orders of magnitude slower than plaintext PyTorch on the text task.
Abstract¶
Secure multi-party computation (MPC) allows parties to perform computations on data while keeping that data private. This capability has great potential for machine-learning applications: it facilitates training of machine-learning models on private data sets owned by different parties, evaluation of one party's private model using another party's private data, etc. Although a range of studies implement machine-learning models via secure MPC, such implementations are not yet mainstream. Adoption of secure MPC is hampered by the absence of flexible software frameworks that "speak the language" of machine-learning researchers and engineers. To foster adoption of secure MPC in machine learning, we present CrypTen: a software framework that exposes popular secure MPC primitives via abstractions that are common in modern machine-learning frameworks, such as tensor computations, automatic differentiation, and modular neural networks. This paper describes the design of CrypTen and measure its performance on state-of-the-art models for text classification, speech recognition, and image classification. Our benchmarks show that CrypTen's GPU support and high-performance communication between (an arbitrary number of) parties allows it to perform efficient private evaluation of modern machine-learning models under a semi-honest threat model. For example, two parties using CrypTen can securely predict phonemes in speech recordings using Wav2Letter faster than real-time. We hope that CrypTen will spur adoption of secure MPC in the machine-learning community.
Source: arxiv
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | MPC tensor-computation library with a PyTorch-compatible API | llm | 0.90 |
| contribution | automatic differentiation and modular neural-network package over secret shares | llm | 0.90 |
| contribution | GPU support and high-performance communication for an arbitrary number of parties | llm | 0.90 |
| contribution | benchmarks on text classification, speech recognition and image classification | llm | 0.90 |
| schemes | secret sharing | llm | 0.90 |
| libraries | CrypTen | llm | 0.90 |
| libraries | PyTorch | llm | 0.90 |
| libraries | Gloo | llm | 0.90 |
| models | linear model over word embeddings | llm | 0.90 |
| models | Wav2Letter | llm | 0.90 |
| models | ResNet | llm | 0.90 |
| models | ViT | llm | 0.90 |
| datasets | Yelp reviews | llm | 0.90 |
| tasks | text classification | llm | 0.90 |
| tasks | speech recognition | llm | 0.90 |
| tasks | image classification | llm | 0.90 |
| tasks | private inference | llm | 0.90 |
| tasks | private training | llm | 0.90 |
| hardware | Intel Skylake 18-core 1.6GHz CPU | llm | 0.90 |
| hardware | NVIDIA P100 GPU | llm | 0.90 |
| threat model | semi-honest, arbitrary number of parties | llm | 0.90 |
| interaction | pure MPC | llm | 0.90 |
| code url | https://github.com/capeprivacy/tf-trusted (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/facebookincubator/gloo (URL found in the full text) | regex | 0.50 |
| code url | https://crypten.ai | llm | 0.90 |
| limitations | fixed-point representation (16-bit precision by default) is prone to overflow/underflow | llm | 0.90 |
| limitations | arithmetic secret shares can suffer wrap-around errors in the 2^64 ring that are hard to debug | llm | 0.90 |
| limitations | no seamless integration with secure data-processing frameworks | llm | 0.90 |
| limitations | differential privacy may be needed to bound leakage from computation outputs | llm | 0.90 |
| limitations | benchmarks run on a single machine, which is faster than a multi-machine deployment | llm | 0.90 |
| llm relevance | medium | llm | 0.90 |
| key result | inference slowdown vs plaintext PyTorch: 2.5-3 orders of magnitude (text-sentiment classification, GPU) | llm | 0.90 |
| key result | inference latency: 0.03 s per sample (text-sentiment classification, two parties) | llm | 0.90 |
| key result | speech recognition speed: faster than real-time (Wav2Letter, two parties) | llm | 0.90 |
| key result | numerical error vs PyTorch: normalized MSE < 4e-4 (text classification) | llm | 0.90 |
| key result | speed-up vs other MPC frameworks: 11-18x vs PySyft, ~3x vs TF-Trusted (MNIST classification on CPU, reported by [35]) | llm | 0.90 |
System card: CrypTen¶
| Group | 5. Interactive MPC and HE+MPC hybrids |
| Family | Secret sharing (PyTorch) |
| Arithmetic | Fixed point |
| Interaction | Interactive multi-party |
| Model scale | BERT (via MPCFormer) |
| Status | Widely used |
| Relation to the SoK | PyTorch-native MPC framework. |
| Note | Used as the backend of MPCFormer and SecFormer. Supports GPUs. |
| Code | https://github.com/facebookresearch/CrypTen |
| Systems page | CrypTen |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: Secret sharing 87 [p.2], MPC 73 [p.1], Differential privacy 6 [p.9], Oblivious transfer 4 [p.3]
- library: PyTorch 30 [p.2], EzPC / CrypTFlow 6 [p.2], TensorFlow 5 [p.2], ABY / ABY3 4 [p.2], CrypTen 4 [p.6], ONNX 2 [p.3]
- model: ResNet 9 [p.4], ViT 7 [p.8], MLP 3 [p.9], Transformer 2 [p.8], AlexNet 1 [p.2], VGG 1 [p.2], SVM 1 [p.10]
- dataset: ImageNet 3 [p.4], Yelp 3 [p.6], MNIST 1 [p.6]
- hardware: GPU 40 [p.1], CUDA 2 [p.5], TPU 1 [p.9]
- technique: ReLU 12 [p.4], Convolution 12 [p.2], Softmax 8 [p.5], Two-party computation 8 [p.2], Matrix multiplication 7 [p.6], Three-party computation 7 [p.2], Malicious model 7 [p.2], Sigmoid 6 [p.5], Semi-honest model 6 [p.1], Non-interactive 4 [p.14], Taylor expansion 3 [p.19], Newton iteration 3 [p.19], Attention 3 [p.8], Quantization 3 [p.20], Polynomial approximation 2 [p.19]
- task: Speech recognition 4 [p.1], Medical / healthcare 3 [p.1]
- topic: Benchmark 17 [p.1], Private / secure inference 4 [p.2], Open source 3 [p.8], Privacy-preserving machine learning 2 [p.10], Homomorphic encryption 2 [p.3]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/5-078_crypten-secure-multi-party-computation-meets-machine-learnin.pdf |
catalogue | 23 | 1.9 MB | 105362d6fa8c |
Database key: crypten-secure-multi-party-computation-meets-machine-learning · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set crypten-secure-multi-party-computation-meets-machine-learning COLUMN VALUE