Pause
Lecture
Moteur de recherche d'offres d'emploi CEA

Can RNN go Homomorphic ? H/F


Vacancy details

General information

Organisation

The French Alternative Energies and Atomic Energy Commission (CEA) is a key player in research, development and innovation in four main areas :
• defence and security,
• nuclear energy (fission and fusion),
• technological research for industry,
• fundamental research in the physical sciences and life sciences.

Drawing on its widely acknowledged expertise, and thanks to its 16000 technicians, engineers, researchers and staff, the CEA actively participates in collaborative projects with a large number of academic and industrial partners.

The CEA is established in ten centers spread throughout France
  

Reference

2021-19245  

Position description

Category

Mathematics, information, scientific, software

Contract

Internship

Job title

Can RNN go Homomorphic ? H/F

Subject

Recurrent Neural Network (RNN) is a subgroup of Neural Network (NN) that creates temporal connections between the NN nodes. It models sequence data for applications such as speech recognition, sentiment classification, DNA sequence analysis and sentences translation. For example, an RNN can be used to translate a French sentence of 15 words to a an English sentence of 10 words.
In this internship, we focus on studying two types of RNN units called Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM). The main idea of GRU and LSTM is to use memory cells to capture the relation with activations from previous RNN layers. GRU and LSTM relies on Sigmoid and Tanh for memory cells and activation computation. One first task consists in evaluating the variation of an RNN accuracy when Sigmoid and Tanh are approximated by the discrete Heaviside and Sign functions. The second task consists in making these units privacy preserving thanks to the use of homomorphic encryption.

Contract duration (months)

6

Job description

Recurrent Neural Network (RNN) is a subgroup of Neural Network (NN) that creates temporal connections between the NN nodes. It models sequence data for applications such as speech recognition, sentiment classification, DNA sequence analysis and sentences translation.
In this internship, we focus on studying two types of RNN units called Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM). The main idea of GRU and LSTM is to use memory cells to capture the relation with activations from previous RNN layers. GRU and LSTM relies on Sigmoid and Tanh for memory cells and activation computation. One first task consists in evaluating the variation of an RNN accuracy when Sigmoid and Tanh are approximated by the discrete Heaviside and Sign functions.
Once GRU and LSTM units are simplified, we move to the second part of the internship where we make this units privacy-preserving. Indeed, as GRU and LSTM can be applied to the analysis of private sequences such as our DNA or our voice, it is interesting to investigate how to make them ensure the confidentiality of the input data during the model evaluation. To do so, we will use homomorphic encryption that allows computation on encrypted data. That is, we will encrypt the inputs to our RNN homomorphically and then we will evaluate GRU and LSTM units on this encrypted data. We will only do RNN classification over encrypted data. Meanwhile, the training will be done on clear data.
Technically speaking, we will have to overcome two challenges. The first challenge is managing the size of the encrypted data to ensure that it does not increase drastically after each unit evaluation while keeping a good accuracy. The second challenge is evaluating the pair (Sigmoid, Tanh) and/or (Heaviside, Sign) on encrypted data.

This internship will take place at CEA, LCYL lab and will last 6 months. It will be supervised by a team with 4 persons with different backgrounds (i.e., crypto, optimization and AI).

Applicant Profile

  • Level: Bac+5 with a major in cryptography, cybresecurity or IA
  • Required skills: basic programming with C/C++ and Python, basic concepts of cryptography, knowledge of NN and RNN is a plus

Position location

Site

Saclay

Job location

France, Ile-de-France, Essonne (91)

Location

  Palaiseau

Candidate criteria

Prepared diploma

Bac+5 - Diplôme École d'ingénieurs

Recommended training

master in crypto, cybersecurity or AI

PhD opportunity

Non

Requester

Position start date

01/01/2022