The AI Framework team is building an open source framework for homomorphic machine learning. The product your team is developing is a compiler which takes the DAG (directed acyclic graph) of a neural network --- or later, of any program --- and transforms it into a fully homomorphic equivalent, with a tunable and small accuracy loss. The main use of the compiler is to convert neural networks into equivalent homomorphic programs operating on end-to-end encrypted data, that can run on insecure servers while maintaining total user privacy. Like the other products of the company, our compiler will eventually go open-source. The main challenges to overcome here are 1) reaching good performances, 2) ensuring a small accuracy loss and 3) not compromising on security.
Zama offers an internship consisting in using our compiler and developing ready-for-use demos of its capabilities. More precisely, the intern will find interesting machine learning applications of neural networks where data privacy is a definite plus and prepare full demos of these neural networks both in the cleartext domain and in their converted form (full self-contained scripts, tutorials, technical explanations, benchmarks). Most likely --- and because building a fully-automated fully homomorphic encryption (FHE) compiler has never been done before ---, this process will end up finding limits or bugs in the compiler: then --- and it may be one of the most exciting parts ---, the intern will be invited to contribute actively to the development of the compiler itself, to fix or find workarounds to overcome the identified issues.
This internship is a unique occasion to work on an experimental tool that has never been done, and to create really exciting demos to show what Zama's technology is capable of. It is also a way to work with some of the best experts in FHE. Ultimately, the candidate will improve her/his knowledge and understanding of machine learning and applied cryptography. Depending on how the internship moves forward, it may be possible to contribute more to the development of the compiler as appropriate.