The Machine Learning team at Zama are building an open source framework for homomorphic machine learning, called the Concrete Framework, gathering tools made by other teams (including the Library team and the Compiler team) with a numpy frontend and ML tools. More precisely, this latter is a set of tools to help our users to train or modify their existing neural networks, such that the obtained neural network is friendly with fully homomorphic encryption (FHE). With our tools the user can develop neural networks which can be used in a context where user's privacy is ensured end-to-end, even if executed on an insecure server.
The goal of our tools is to be as easy and user-friendly as possible, while preserving the accuracy of networks as much as possible. Notably, we don't want to require the user to understand anything about cryptography, or, said in another way, we want to provide tools which look similar to what data scientists are already using day to day.
The proposed internship is of an internal hacker style: the intern will use the Concrete Framework and build on top of it shiny demos and usecases, covering different aspects of machine learning. At the end, the best usecases will be added to a public repository. If ever new efficient techniques are discovered, patents and publications may be considered.
We are looking for an intern with enough familiarity with ML and ideally with some knowledge on quantized / discretized / binary networks. The day to day activity will be about understanding and using the framework, and building usecases with it. Possibly, if some bugs are found, if interested, the intern might be included in the SW team to help fixing them. Understanding of cryptography is a plus, but not required.
As an intern, with our guidance, you will be responsible for:
- training neural networks to follow some constraints of FHE networks
- post-modifying trained networks to follow some constraints of FHE networks
- using the Concrete Framework to compile these networks into FHE equivalents, and report possible bugs or difficulties
- publishing in a repository the best usecases
- if enough results, co-writing patents
- if enough results, co-writing papers in research conferences