Giulio Rossolini

Research profile

Research summary

Giulio Rossolini is a researcher at Scuola Superiore Sant'Anna (Italy), name-resolved here from public bibliographic data (OpenAlex identity, ORCID 0000-0002-6404-2627). The record shows a strongly security-oriented machine learning portfolio: topic counts are led by Adversarial Robustness in Machine Learning (25), followed by Security and Verification in Computing (13), Topic Modeling (10), Digital Media Forensic Detection (10), and Advanced Neural Network Applications (9). Output has accelerated sharply — yearly counts rise from 1 (2020) and 4 (2022) to 8 (2023), 10 (2025), and 16 (2026), with 28 works in the last three years — while indicators on file are h-index 8 and i10-index 8. Two 2026 arXiv papers reviewed by Evolve illustrate the range: DANTINOX, a JAX/Flax framework unifying autoregressive decoding, masked discrete diffusion, and flow matching on one Transformer backbone, and an attack study showing that token shuffling in ViT split inference provides only apparent privacy. Both carry 'high' reproducibility readiness; the reproducibility mean on file is 100, computed over 1 evaluated paper. On both reviewed papers he holds a middle, non-corresponding author position within a compact circle of 5 distinct coauthors. Caveats apply: the output and impact radar axes are not yet scored.

Research keywords

Highlights

• DANTINOX unifies autoregressive, diffusion, and flow-matching language modeling

  Evolve's review finds a credible JAX/Flax framework that unifies autoregressive decoding, masked discrete diffusion, and flow matching on a single Transformer backbone, with solid engineering and honestly scoped claims; the small-scale, undertrained comparisons are suggestive rather than definitive. Reproducibility readiness is rated high.

• Token shuffling in ViT split inference offers only apparent privacy

  The reviewed study presents a feature-inversion attack (SARA) against Vision Transformer split inference that uses token reduction and shuffling, showing the defense's privacy is largely illusory. Evidence is convincing on ImageNet with ViT-B/16, though the scope is narrow and no error bars are reported. Reproducibility readiness is rated high.

• Publication tempo accelerating around an adversarial-robustness core

  Recorded yearly counts rise from 1 (2020) to 16 (2026), with 28 works in the last three years, while topic counts are led by Adversarial Robustness in Machine Learning (25), ahead of Security and Verification in Computing (13) — a steepening, security-centered research agenda.

Timeline

1. 2020First recorded publication year in the snapshot, with 1 work counted.
2. 2023Output climbs to 8 works in the year, the first clear step in the ramp-up.
3. 2024A dip year, with only 2 works recorded.
4. 2025Output rebounds to 10 works in the year.
5. 2026Peak recorded year: 16 works, including both Evolve-reviewed papers — the DANTINOX framework and the ViT split-inference attack study.

Papers in the Evolve catalog

• 2026-09-17

  DANTINOX: A Unified Framework for Multi-Paradigm Language Modeling

• 2026-09-01

  Position Matters: Feature Inversion Attacks in ViT Split Inference with Token Reduction and Shuffling

Frequent co-authors