Two papers by CAIM teams at ICML 2026 ranging from embedding approaches to adversarial robustness.
Riemannian Metric Matching
Bamberger, J., Gosztolai, A., Vendergheynst, P. Bronstein, M., Jones, I. Riemannian Metric Matching for Scalable Geometric Modelling of Distributions, Proc. ICML 2026. paper, ICML 2026
This work introduces Riemannian metric matching, a neural framework for learning the underlying geometry of high-dimensional data without relying on costly graph or kernel constructions. By estimating geometric structure directly from perturbed samples, the method enables scalable, graph-free analysis of complex datasets such as images, while retaining theoretical consistency and competitive accuracy against classical diffusion geometry approaches.
Adversarial Robustness
Dhar, J., Pandey, M., Bozorgtabar, B., Zaidi, N., Zhang, W., Li, W., Mu, T., Mahapatra, D., Baktashmotlagh, M., Le, T., Chen, C., Mistry, S., Gonzalez, C., Ebrahimi Kahou, S., Yao, L., Koniusz, P., Fisher, R., Phung, D., Han, B., Vasconcelos, N., & Lió, P. (2026). Does a hybrid space-aware randomized defense improve empirical and certified adversarial robustness? Proceedings of the 43rd International Conference on Machine Learning (ICML 2026).
ICML 2026
HySCAN is a hybrid randomized defense that strengthens imaging models by combining stochasticity in the network weights with noise injection in intermediate feature representations. This dual strategy helps bridge the gap between formally certified robustness and practical robustness against strong adversarial attacks, while preserving clean accuracy. Across natural and medical imaging datasets, including ImageNet, HAM10000, and NIH Chest X-ray, HySCAN outperforms existing certified and empirical defenses.