One of the most important uses of encrypted machine learning technology is the ability for an AI model to be trained simultaneously across multiple encrypted datasets that may be owned by different organizations. Doing so provides the highest level of security to protect not only the datasets themselves but also the model/statistics being derived from the data. However, the technical expertise and computational complexity involved in providing this level of protection create a very high barrier to entry. The goal of this project is to make an application (with a UI) that lowers the barrier to entry for the encrypted training of linear models. In particular, a genetics researcher should be able to upload a CSV of datapoints into a server they own which will coordinate with other servers to train an encrypted linear model. This UI should be deployable using a Grid server.
In this project, I will be working with Jon Bloom of the Broad Institute of Harvard and MIT for roadmap guidance and model development. In particular, Jon has already articulated a specific training regime for model training, motivated by his work on the open-source Hail project for scalable genomic analysis.
- Scalable Genomic Analysis
- Jon Bloom: Homepage, GitHub
- OpenMined: Homepage, GitHub