Ongoing Projects
My group builds rigorous theory and principled algorithms for trustworthy machine learning; proofs first, then the algorithms and open-source tools those proofs justify. Two programmes anchor the work. I am actively looking for PhD students and postdocs who are excited by provable guarantees and clean algorithm design; if either thread speaks to you, write to me at nigu[at]di.ku.dk.
ATLAS Adversary-Tailored Learning Scheme
Making privacy and robustness affordable by calibrating them to realistic adversaries rather than worst-case ones.
Differential privacy and Byzantine resilience are the gold-standard guarantees for private, robust learning, yet both are calibrated to an omniscient, all-powerful adversary, and the protective noise this demands often collapses accuracy. ATLAS starts from one observation: privacy and robustness are the same sensitivity problem seen through two adversary lenses. We replace worst-case sensitivity with expected adversarial sensitivity (EAS), the sensitivity to a realistic, resource-bounded adversary, and cast the defender's task as an online-learning problem, so smart defences adapt to the adversary's observed behaviour. This yields auditable certificates and training algorithms that recover most of the accuracy today's tools give away.
You might work on: generalisation bounds under differential privacy, online and adaptive defences against bounded adversaries, reconstruction-risk certification, or the open-source EAS toolkit.
AGORA Across-architecture Generalization via Open Representation Alignment
Letting independently trained models exchange knowledge across different architectures, with provable guarantees.
Federated learning assumes every participant shares one architecture, so it can only average parameters. AGORA drops that assumption: models living in different hypothesis spaces, say a lightweight convolutional net, a large transformer, or a physics-informed model, exchange knowledge through their predictions rather than their weights, via functional-distance alignment with PAC-Bayes guarantees. The aim is open infrastructure for collaborative learning that no single institution owns, a step toward developing large-scale AI models democratically rather than concentrating them in a few hands, with robustness, privacy, unlearning and fairness built in from the start.
You might work on: functional distillation and representation alignment (optimal-transport / Bregman geometry), generalisation theory for cross-architecture merging, or robust and privacy-preserving aggregation protocols.
A few related threads feed into both programmes: robustness in decentralised learning over sparse communication graphs, robust federated inference, and privacy-preserving collaboration for healthcare. Background reading: my book on Robust Machine Learning and my chapter in Large Language Models in Cybersecurity on the difficulty of provable robustness and privacy for LLMs.