LaRiffle

Collateral Learning

Motivation

Imagine that you train a neural network to perform a specific task, and you discover it has also learned to do another completely different task, which is very sensitive. Is this possible? What can you do to prevent this?

Context

Let's assume you have a semi-private trained network performing a prediction task of some nature pred1, ie a network with the first layers encrypted and the last ones visible in clear. The pipeline of the network could be written like this: x -> Private -> output(x) -> Public -> pred1(x). pred1(x) could be the age based on an face picture input x, or the text traduction of some speech record.

LaRiffle

The SQL Assistant

n-gram embedding

We use embedding in vectors but are aware of limitations

Build and select n-grams

We won't embed words as in our context we have very unfrequent words (such as addresses, phone numbers, family names, etc) but n-grams. We consider bi, tri and tetra-grams, made of letters, numbers, regexp masks (\d, etc). As we might have numerous n-grams, we will build a score that mixes frequency and discriminacy (with entropy feature for instance), and keep the N elements with the highest score.

LaRiffle

Get ready to start 🚀

Before starting working, you should make sure you have all the tools you need!

Python

Python is the programming language we will use for the algorithmic part of the project. Open a terminal (or cmd.exe on Windows) and type:

python --version