Today's Internet can be seen as a huge data store that collects personal and sensitive data about its users. This leads to significant security and privacy risks for end users, who lose control over the data they share. Developing methods and tools to enable a secure and privacy-friendly processing of data thus constitutes a core challenge to all data-driven ecosystems and applications. In particular, the success of digitalization heavily depends on whether companies are able to gain their users' trust regarding the protection of their privacy. This research area strives to develop disruptive new frameworks for reasoning about and improving security and privacy in information processing in various settings, efficiently and at scale. In the last years, this area had a particular focus on the following topics: novel methods and tools for the algorithmic sanitization of privacy-sensitive data, in particular for genomic and medical research; new techniques for quantitatively assessing end user privacy; as well as efficient techniques for secure, verifiable computation.
Conference on Neural Information Processing Systems (NeurIPS)
Differentially Private Bilevel Optimization: Efficient Algorithms with Near-Optimal Rates
International Conference on Machine Learning (ICML)
Enhancing LLM Training via Spectral Clipping
Annual Meeting of the Association for Computational Linguistics (ACL)
Conference on Learning Theory (COLT)
On the Stability of Nonlinear Dynamics in GD and SGD: Beyond Quadratic Potentials
IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
Bridging Domains through Subspace-Aware Model Merging
