Sepehr Khodadadi | MaLGa (Machine Learning Genoa Center), University of Genoa

About Me

I am an M.Sc. student in Data Science and Engineering (AI track) at the University of Genoa, with a 30.2/30 GPA (lode counted as 31). I work at the intersection of machine unlearning, data privacy, and trustworthy machine learning at MaLGa (Machine Learning Genoa Center), in collaboration with Télécom Paris, advised by Prof. Vito Paolo Pastore and Prof. Enzo Tartaglione.

I am driven by one goal: making trained models controllable, so that what they memorize and expose can be measured, bounded, and removed, while making those guarantees verifiable. For my thesis, Machine Unlearning for Model Debiasing, I implemented unlearning methods and evaluated them under evolving criteria of this literature. That work led to TRAIL, an open-source evaluation library and my first first-author paper:).

Research Interests

Privacy-preserving machine learning: machine unlearning, differential privacy, federated learning, and synthetic data; model evaluation and privacy auditing; deep learning for computer vision and biomedical imaging, with a focus on semantic segmentation, representation learning, and 3D reconstruction.

News

[Aug. 2026] Our paper TRAIL: Navigating the Jungle of Machine Unlearning Evaluation is accepted as an oral presentation at the ECCV 2026 Workshop and Challenge on Unlearning and Model Editing (U&ME), my first accepted publication.
[Jul. 2026] First peer-review assignment: serving as a reviewer for the ECCV 2026 FAILED Workshop (Model Debiasing).
[Jul. 2026] Released ALoG, a training-free, interpretable-by-design pipeline for retinal vessel segmentation on the DRIVE dataset.
[Feb. 2026] Began my M.Sc. thesis research on machine unlearning at MaLGa, co-supervised with Télécom Paris.

Publications

TRAIL: Navigating the Jungle of Machine Unlearning Evaluation

S. Khodadadi Hosseinabadi, E. Iakovleva, E. Tartaglione, and V. P. Pastore

Accepted (oral presentation) at the ECCV 2026 Workshop and Challenge on Unlearning and Model Editing (U&ME), 2026.

OpenReview Code Oral presentation

Application of the Brain Drain Optimization Algorithm to the N-Queens Problem

S. Ramezani Jolfaei and S. Khodadadi Hossein Abadi

arXiv preprint arXiv:2504.18953, 2025.

PDF GScholar

Projects

TRAIL (machine-unlearning evaluation library): reproducible, protocol-only evaluation of unlearned checkpoints; the library behind the U&ME workshop paper.
Multi-agent clinical decision support: BDI multi-agent system fusing chest X-ray imaging agents and report-text NLP agents under a Prolog consensus engine (graduate course project; not for clinical use).
ALoG (interpretable retinal vessel segmentation): training-free, interpretable-by-design classical pipeline on DRIVE.
Attention U-Net for retinal vessel segmentation: hybrid Dice+Focal loss after densely augmenting 20 images into 63,800 patches.
3D geological modeling: rock-mass slicing, optimal block extraction, and volumetric analysis (Trimesh, Open3D).
Shifting Spotlights (media-bias visualization): D3.js scrollytelling on global conflicts and media attention (live demo).
AR gallery presenter: Unity/Vuforia app that anchors a 3-D presenter beneath recognized paintings.