Skip to content

Programmable Bootstrapping Enables Efficient Homomorphic Inference of Deep Neural Networks

Authors: Chillotti, Joye, Paillier
Year: 2021
Venue: CSCML 2021

type: scheme collection: catalogue, schemes status: auto

Links: ePrint 2021/91 · link

Summary

Auto-extracted summary (unreviewed)

Zama's paper introduces programmable bootstrapping (PBS) for a TFHE variant: the blind-rotation test polynomial is programmed as a look-up table so that any function of an encrypted value is evaluated while refreshing noise. Combined with encoding methods, this yields a general framework for homomorphic inference of deep neural networks that scales with the number of layers, needs no model retraining, and is implemented in the open-source Rust library Concrete; the authors report first experiments at 80- and 128-bit security.

Abstract

In many cases, machine learning and privacy are perceived to be at odds. Privacy concerns are especially relevant when the involved data are sensitive. This paper deals with the privacy-preserving inference of deep neural networks. We report on first experiments with a new library implementing a variant of the TFHE fully homomorphic encryption scheme. The underlying key technology is the programmable bootstrapping. It enables the homomorphic evaluation of any function of a ciphertext, with a controlled level of noise. Our results indicate for the first time that deep neural networks are now within the reach of fully homomorphic encryption. Importantly, in contrast to prior works, our framework does not necessitate re-training the model.
Source: pdf

Extracted details

Field Value Source Confidence
contribution programmable bootstrapping: look-up-table evaluation of arbitrary functions during TFHE bootstrapping llm 0.65
contribution general framework for homomorphic evaluation of deep neural networks (dense, conv and non-linear layers) without retraining llm 0.65
contribution new Rust library (Concrete) implementing the extended TFHE variant llm 0.65
contribution first experiments indicating deep networks are within reach of FHE llm 0.65
schemes TFHE llm 0.65
schemes GLWE/LWE llm 0.65
schemes GGSW llm 0.65
libraries Concrete llm 0.65
models deep neural networks llm 0.65
datasets MNIST llm 0.65
tasks private neural network inference llm 0.65
tasks image classification llm 0.65
security level 80-bit and 128-bit llm 0.65
threat model untrusted server, client-server llm 0.65
interaction non-interactive FHE llm 0.65
code url https://bitbucket.org/malb/lwe-estimator/ (URL found in the full text) regex 0.50
code url https://github.com/zama-ai/concrete.Acknowledgments (URL found in the full text) regex 0.50
code url https://github.com/zama-ai/concrete llm 0.65
limitations impact of specialised hardware not yet studied (several orders of magnitude could be gained) llm 0.65
limitations homomorphic training of neural networks not addressed llm 0.65
llm relevance medium llm 0.65

System card: TFHE programmable bootstrapping for DNNs

Group 4. Alternative FHE schemes
Family TFHE PBS
Arithmetic Quantized integers
Interaction Non-interactive
Model scale Deep quantized NNs
Status Active
Relation to the SoK Evaluates activations exactly with lookup tables.
Note Basis of Zama Concrete ML, which supports quantized LLM fragments such as GPT-2 attention heads.
Code https://github.com/zama-ai/concrete-ml
Systems page TFHE programmable bootstrapping for DNNs

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: TFHE 12 [p.1], GSW 4 [p.2], Paillier 2 [p.1], Leveled / somewhat HE 2 [p.1]
  • library: Zama 2 [p.1], ONNX 2 [p.11]
  • model: Logistic regression 1 [p.13]
  • dataset: MNIST 3 [p.2]
  • hardware: NVIDIA A100 1 [p.11], Intel Xeon 1 [p.11], GPU 1 [p.11]
  • technique: Bootstrapping 41 [p.1], Programmable bootstrapping 17 [p.1], Rotations 8 [p.7], Key switching 7 [p.8], ReLU 7 [p.2], Convolution 6 [p.9], Lookup table 5 [p.7]
  • assumption: LWE 32 [p.3], Lattice-based 5 [p.13]
  • task: Language modeling / generation 4 [p.4], Genomics 3 [p.12]
  • topic: Homomorphic encryption 27 [p.1], Open source 2 [p.12]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/4-055_programmable-bootstrapping-enables-efficient-homomorphic-inf.pdf catalogue 14 487.8 KB 17ef52d3a8e2
references/schemes/s22-1_eprint-2021-091.pdf schemes 14 487.8 KB 17ef52d3a8e2 byte-identical to references/catalogue/4-055_programmable-bootstrapping-enables-efficient-homomorphic-inf.pdf

Database key: programmable-bootstrapping-enables-efficient-homomorphic-inference-of · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set programmable-bootstrapping-enables-efficient-homomorphic-inference-of COLUMN VALUE