Building blocks for LSTM homomorphic evaluation with TFHE - CEA - Commissariat à l’énergie atomique et aux énergies alternatives Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Building blocks for LSTM homomorphic evaluation with TFHE

Résumé

Long Short-Term Memory (LSTM) is a Neural Network (NN) type that creates temporal connections between its nodes. It models sequence data for applications such as speech recognition, image captioning, DNA sequence analysis, and sentence translation. Applications that are often subject to privacy constraints. This paper thus presents basic building blocks for the homomorphic execution of an LSTM that would respect the privacy of its inputs. By means of TFHE functional bootstrapping, we propose several approaches for homomorphically evaluating discretized flavors of the Sigmoid and Tanh activation functions. Experimental results show that the accuracy of the resulting discretized networks remains comparable to a full precision clear-domain execution. Performance-wise, we are able to homomorphically compute a Sigmoid or Tanh function in 0.3 or 0.15 secs (depending on whether or not multivalue bootstrapping is relied on). This paves the way towards evaluating practical LSTMs over encrypted inputs in around 1 to 3 mins which is competitive with the state of the art.
Fichier principal
Vignette du fichier
CSCML.pdf (567.79 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

cea-04186110 , version 1 (23-08-2023)

Identifiants

  • HAL Id : cea-04186110 , version 1

Citer

Daphne Trama, Pierre-Emmanuel Clet, Aymen Boudguiga, Renaud Sirdey. Building blocks for LSTM homomorphic evaluation with TFHE. 7th International Symposium on Cyber Security, Cryptology and Machine Learning (CSCML 2023), Jun 2023, Paris, France. pp.117. ⟨cea-04186110⟩
37 Consultations
51 Téléchargements

Partager

Gmail Facebook X LinkedIn More