His research is centered on two complementary directions. On one side, he develops Graph Neural Networks and geometric deep learning models for bioinformatics and molecular data analysis. His work includes the design of advanced architectures such as Layered GNNs (LGNNs) and Composite GNNs for heterogeneous graphs, applied to problems like protein–protein interface prediction, drug side-effect modeling, ribosome translation speed analysis, and health effect prediction from complex biological networks. He has also contributed to the development of GNNkeras, a high-level framework for graph-based deep learning.
On the other side, his research addresses robustness and security of deep learning systems, with a focus on adversarial machine learning and synthetic media. He proposed the Jacobian-induced Mahalanobis distance Attack (JMA), a theoretically grounded method for crafting near-optimal targeted adversarial examples, particularly effective in complex and multi-label settings. His work on deepfake detection explores transfer learning and semantic segmentation strategies for identifying synthetic faces, as well as the analysis of model vulnerabilities—showing how many detectors rely on spurious background cues rather than meaningful facial features.
He has been actively involved in several research projects bridging these domains. In the Tuscany Health Ecosystem (THE) project, he contributed to AI methods for bioinformatics, including molecular modeling and protein interaction prediction using GNNs. In the SOS-AI / AI-RESCUE project, he is working on adversarial scenarios for synthetic media detection, evaluating both attack and defense strategies under realistic threat models.
His research interests include geometric deep learning, computational bioinformatics, multimedia forensics, AI security, and the development of robust and interpretable machine learning models for real-world applications.
