Below is a list of some featured publications. For a complete list, check out Google Scholar.
Analyzed the sensitivity of image-to-image dehazing transformers w.r.t. adversarial attacks, and developed empirical defenses to increase their robustness. 
Introduced a class of neural-networks (i.e. ABBA) which can approximate any feed-forward structure using only positive weights. For this class, tight Lipschitz bounds can be computed and imposed, through which we trained robust networks that outperform standard AT at high perturbations. 
Developed a multi-objective training mechanism focused on balancing the trade-off between good reconstruction and new details when super-resolving Sentinel-2 images. Seen as an inverse problem, super-resolution can be pushed either towards higher synthesis or higher consistency. 
Part of my BSc thesis, designed an architecture for lesion segmentation from CT scans of lungs. 