PhD Position in Differentially Private Foundation Models

We are seeking a highly motivated PhD candidate to explore differentially private machine learning methods for foundation models including LLMs. Examples of possible research topics include developing new theoretical analyses and algorithms that enhance the privacy-protection in foundation models [1,2], creating robust and theoretically sound empirical methods to evaluate privacy-preserving properties of the models [3], and advancing zero-shot and few-shot learning capabilities under privacy constraints [4] including retrieval augmented generation type of methods. This position offers a unique opportunity to contribute to understanding and practical applications of privacy-preserving AI.

The candidate would ideally have a strong background in machine learning and mathematics, with an interest in both theoretical development and practical implementation.

This research project is a collaboration between the University of Helsinki and Nokia Bell Labs, and the student will have as mentors Prof. Antti Honkela (UH) and Dr Antti Koskela (Nokia Bell Labs).

[1] Koskela, A., Jälkö, J., Prediger, L., & Honkela, A. Tight differential privacy for discrete-valued mechanisms and for the subsampled Gaussian mechanism using FFT. In International Conference on Artificial Intelligence and Statistics (2021).
[2] Koskela, A., Redberg, R., & Wang, Y.-X.. Privacy Profiles for Private Selection. Forty-first International Conference on Machine Learning (2024).
[3] Tobaben, M., Pradhan, G., He, Y., Jälkö, J., & Honkela, A. Understanding Practical Membership Privacy of Deep Learning. In Privacy Regulation and Protection in Machine Learning (ICLR 2024 workshop).
[4] Tobaben, M., Shysheya, A., Bronskill, J., ... & Honkela, A. On the Efficacy of Differentially Private Few-shot Image Classification. Transactions on Machine Learning Research (2023).

Why Helsinki:

• Vibrant AI research community with strong links to leading researchers in Europe (e.g. ELLIS, ELSA)
• Access to very good computational resources, including #3 fastest supercomputer in the world, LUMI
• Very livable city, very easy to get by using English

How to apply:

Application instructions (DL 9 September 2024)