Topics in trustworthy machine learning and AI: Robustness, privacy and data heterogeneity

Machine learning and statistical algorithms are now implemented at a large scale in almost every aspect of our society, significantly impacting our daily lives through their performance. Hence, there is a soaring demand for the development of trustworthy procedures. Two projects are available under the broad theme of Topics in trustworthy machine learning and AI.

(1) Differential privacy of sampling algorithms. Sampling algorithms inherently possess privacy properties due to its probabilistic nature. Although such nature has been explored for simple sampling algorithms, there is considerable room for further investigation into more sophisticated sampling schemes, the effects of subsampling, and relaxations of the conditions on the target distribution.

(2) Robust and private learning in heterogenous and distributed environments. Building on previous work in this area, we plan to tackle more challenging data types, such as network and tensor data, and to study the effects of data heterogeneity, privacy, and contamination within a unified framework.

References:
Li, M., Berrett, T.B. and Yu, Y., 2023. On robustness and local differential privacy. The Annals of Statistics, 51(2), pp.717-737.

Li, M., Tian, Y., Feng, Y. and Yu, Y., 2024. Federated transfer learning with differential privacy. arXiv preprint arXiv:2403.11343.

Bertazzi, A., Johnston, T., Roberts, G.O. and Durmus, A., 2025. Differential privacy guarantees of Markov chain Monte Carlo algorithms. arXiv preprint arXiv:2502.17150.