Trustworthy & Private Machine Learning
Developing federated learning and differential privacy techniques that let sensitive data — especially in healthcare — be used for training without compromising individual privacy.
Why this matters
Healthcare, financial, and behavioral data are some of the most useful sources for training machine learning models — and some of the most sensitive. Centralizing this data for training is often legally or ethically infeasible, which makes privacy-preserving collaboration a practical necessity rather than a nice-to-have.
What we work on
- Federated learning protocols that reduce communication overhead while tolerating unreliable, non-IID clients.
- Differential privacy mechanisms with tight, empirically validated utility bounds.
- Auditing tools that quantify what a trained model actually leaks about individual training examples.
Representative outcomes
This line of work underlies the FedCare project (see Current Projects) and several publications on federated clinical NLP, including formal privacy analyses that have been adopted by clinical collaborators evaluating their own data-sharing agreements.
Back to all research
Abhiraj Bibhar
Department of Sociology and Anthropology
SRM University - Andhra Pradesh