Efficient models
for real data.

Sapienza University of Rome · May 2026 to Present · with Paolo Russo

Physically Grounded SSMs for Hyperspectral Super-Resolution

Proposed AbSpec-Mamba and VolMamba v2, models that condition selective state transitions on per-pixel material abundance, using prototype-guided reconstruction to recover fine spectral detail without hallucinating physically implausible spectral signatures.

Building content-adaptive multidimensional scanning and progressive coarse-to-fine reconstruction so the model restores high-resolution hyperspectral images under severe sensor degradation and spectral corruption.

University of Tübingen · Sep 2025 to Present · with Andreas Ziegler

CEGNN: Rate as a Feature, Not a Nuisance

Event-camera output rate depends on scene motion, texture and sensor settings. On N-Caltech101 it spans roughly 26× across recordings and differs by class. Most event-GNN pipelines either rescale timestamps, discarding the rate before the network sees it, or train for rate invariance and lose accuracy on real recordings.

CEGNN keeps the rate and uses it: the log event rate drives identity-initialised FiLM layers in a PointNet++ backbone, trained with a two-stage rate curriculum. +14.0 points on DVS128 Gesture and +6.3 on N-Caltech101 at natural rates, and +34 points on N-Caltech101 under synthetic 1/8× to 8× speed shifts, for 49K extra parameters. Negative results included. arXiv → Code →

IIT Ropar · May to Aug 2025 · with Dr. Puneet Goyal

SCAF: Robust Hyperspectral Classification via Optimal Transport

Identified that standard HSI pipelines discard recoverable inter-band correlation by treating bands independently. Recovered this signal using differentiable optimal-transport grouping with balanced Sinkhorn assignments.

Per-group Mamba processors isolate sensor noise to individual groups, converting spectral redundancy into a graceful-degradation mechanism, achieving +11.6% accuracy under 20 to 30% band dropout against Transformer baselines.

NIT Kurukshetra · Sep 2024 to May 2025 · with Dr. Vishwas Rathi & Dr. Shweta Sharma

3D Medical Segmentation & Malware Analysis

Built parameter-efficient 3D segmentation models targeting deployability on memory-constrained hardware, work that became LightMedSeg and RefineFormer3D. Concurrently developed hybrid frameworks for ransomware detection combining static binary features with NLP-based code representations, leading to the ICDAM 2025 paper.

Machine unlearning · 2026 · PACE with Paolo Russo

Privacy Leakage in Class Unlearning: ADVR & PACE

Segmentation unlearning that merely relabels a class leaves a class-shaped hole in the model's output, and a learned attacker can reconstruct the forgotten mask from the output alone. ADVR (adversarial silhouette removal plus retain replay) is the only method leak-safe on both output and feature surfaces across ADE20K and Cityscapes, while keeping more retained-class accuracy than the baselines. Code →

PACE extends the audit to video: averaging a model's output over a ~2-second clip recovers the forgotten class from every published baseline tested, and a time-shuffled control shows that cross-frame redundancy, not trajectory, carries the signal. A permutation-invariant pooled adversary cuts pooled attacker AP from 93.8 to 18.8 on VSPW (person). arXiv → Project page →