Privacy preserving ML and neuroscience

Overview:

We will primarily pursue and extend our previous work on privacy preserving machine learning (PPML). We will use split learning methods, but are open to using appropriate aspects of Federated learning also.

Machine learning for health care, including C19 diagnosis, vaccination and associated treatments, contact tracing and checking of paths for C19 presence, sanitation, etc, faces the problem of medical ethics and confidentiality of data, in a typically distributed setting, for which PPML is one approach. Hospitals and health care centers do not share data about their patients, limiting the applicability of ML techniques, due to small fragmented datasets. PPML addresses this issue.

We have worked with MIT on this aspect, and published two papers in NeurIPS2019 workshops, on privacy preserving machine learning – split learning (an alternative to Federated learning), and out-of-class detection – urine analysis. In parallel work we have almost achieved Kaggle Gold status in fundus retinopathy.

We will extend MIT’s work in PPML, and contact tracing/ identification of containment zones (CovidSafePaths renamed as PathCheck www.pathcheck.org ) as follows.

• We will design and implement a scalable infrastructure for PPML, easily deployed.
• We will extend PPML to unlabeled data, which will be especially useful for monitoring the pandemic in rural ares with 70-80 Crore people, where only gross symptoms can be observed with available equipment, and RT-PCR tests will be accessible only to a few.
  • It is especially important to diagnose infection by mutants, where the symptoms are not the same as the normal variety, and only extensive testing (RT-PCR), accessible only to a few, can properly do labelling. Hence machine learning techniques with at best limited labelling has to be used, and ideally with privacy.
• The extensions can be also applied for monitoring vaccine side effects, without determining whether a vaccinated person has been re-infected or not.