Oğuzhan Ersoy

Hello!

Ozzy is here…

I am a researcher who tries to apply cool Math formulas to real-life problems, and I like challenging myself with new technologies. Currently, my work sits on the security and efficiency of decentralised and collaborative AI — how to train, fine-tune and verify models across machines that do not trust each other.

I am the Research Lead at Gensyn, which I joined in 2023. My research there is on making decentralised training and inference hold up in practice: staying efficient when the machines are heterogeneous and unreliable, and staying secure when the participants cannot be assumed honest. Much of that is about verifiability: being able to check that a training run or an inference really happened as claimed.

Before Gensyn I was a postdoc at Radboud University (2022–23), working on the security of AI: robust collaborative machine learning, and in particular backdoor and poisoning attacks and the defences against them. I supervised PhD students and taught master’s courses there. Earlier I was at TU Delft — a postdoc in 2021–22, and a graduate researcher from 2017 to 2021 — where I worked on the security, privacy, scalability and economic aspects of blockchain technology. Going from blockchain protocols to decentralised model training turned out to be a smaller jump than it sounds: both are about getting parties who do not trust each other to agree on what actually happened.

On the education side, I did my Ph.D. in the Cyber Security Group at TU Delft. Before Delft I was at Bogazici University in Istanbul, first for a double-major B.Sc. in Electrical & Electronics Engineering and Mathematics, then for an M.Sc. in Electrical & Electronics Engineering.

Personally, I would define myself as a sporty person. I like to try and fail at new things. Lately, I am learning (and performing) improvised theather. I play amateur-level chess but also enjoy watching Super GM tournaments. To keep my creative side alive, I also paint once in a while.

news

Oct 5, 2026	Attended Dagstuhl Seminar 26401, Divided We Stand: Security and Privacy of Collaborative Learning — a great week of exchanging ideas on the open challenges in the field.
Sep 28, 2026	Our paper with Eros Fanì, Training-Free Dynamic Upcycling of Expert Language Models, is accepted at NeurIPS 2026! We introduce DUME, a way to combine already-trained domain experts into one Mixture-of-Experts model without post-merge training.
Sep 15, 2026	OPEN-1B is out — a fully auditable LLM whose training steps can be replayed and verified bit for bit. Preprint · announcement.
Sep 8, 2026	We built IR3DE-AXL, a collective inference network prototype where participants can contribute expert models, use models served by others, or both.
Jun 6, 2026	IR3DE: A Linear Router for Large Language Models is accepted at the ICML 2026 Workshop AdaptFM. Preprint · code.
May 22, 2026	Wrote up F-TIS on the Gensyn blog — running GRPO across heterogeneous models without assuming they are all the same.

selected publications

NeurIPS

Training-Free Dynamic Upcycling of Expert Language Models

Eros Fanı̀, and Oğuzhan Ersoy

In Advances in Neural Information Processing Systems (NeurIPS) 2026

Machine Learning Model Merging

arXiv Bib

@inproceedings{fani2026upcycling,
  title = {Training-Free Dynamic Upcycling of Expert Language Models},
  author = {Fan{\`\i}, Eros and Ersoy, O\u{g}uzhan},
  booktitle = {Advances in Neural Information Processing Systems ({NeurIPS})},
  year = {2026},
}

arXiv

OPEN-1B: A Fully Auditable Training Run

John Donaghy, Brian Wilcox, Oğuzhan Ersoy, Shikhar Rastogi, Adam St Arnaud, Alexey Titov, Jordan Greenberg, Ben Fielding, and Harry Grieve

arXiv preprint arXiv:2609.17380 2026

Machine Learning Verifiability Pre-Training

arXiv Bib

@article{donaghy2026open1b,
  title = {{OPEN-1B}: A Fully Auditable Training Run},
  author = {Donaghy, John and Wilcox, Brian and Ersoy, O\u{g}uzhan and Rastogi, Shikhar and St Arnaud, Adam and Titov, Alexey and Greenberg, Jordan and Fielding, Ben and Grieve, Harry},
  journal = {arXiv preprint arXiv:2609.17380},
  year = {2026},
}

ACL-F

Hail to the Thief: Exploring Attacks and Defenses in Decentralised GRPO

Nikolay Blagoev, Oğuzhan Ersoy, and Lydia Yiyu Chen

In Findings of the Association for Computational Linguistics: ACL 2026

Machine Learning Security and Privacy Decentralisation Post-Training

arXiv Bib

@inproceedings{blagoev2026grpo,
  title = {Hail to the Thief: Exploring Attacks and Defenses in Decentralised {GRPO}},
  author = {Blagoev, Nikolay and Ersoy, O\u{g}uzhan and Chen, Lydia Yiyu},
  booktitle = {Findings of the Association for Computational Linguistics: {ACL}},
  pages = {39130--39148},
  year = {2026},
}

ICLR-W

Backdoor Attacks on Decentralised Post-Training

Oğuzhan Ersoy, Nikolay Blagoev, Jona Lintelo, Stefanos Koffas, Marina Krček, and Stjepan Picek

In ICLR Workshop on Principled Design for Trustworthy AI 2026

Machine Learning Security and Privacy Decentralisation Post-Training

arXiv Bib

@inproceedings{ersoy2026backdoor,
  title = {Backdoor Attacks on Decentralised Post-Training},
  author = {Ersoy, O\u{g}uzhan and Blagoev, Nikolay and te Lintelo, Jona and Koffas, Stefanos and Kr{\v{c}}ek, Marina and Picek, Stjepan},
  booktitle = {{ICLR} Workshop on Principled Design for Trustworthy {AI}},
  year = {2026},
}

arXiv

NoLoCo: No-all-reduce Low Communication Training Method for Large Models

Jari Kolehmainen, Nikolay Blagoev, Semih Kara, John Donaghy, Christopher Nies, and Oğuzhan Ersoy

arXiv preprint arXiv:2506.10911 2025

Machine Learning Decentralisation Pre-Training

arXiv Bib

@article{kolehmainen2025noloco,
  title = {{NoLoCo}: No-all-reduce Low Communication Training Method for Large Models},
  author = {Kolehmainen, Jari and Blagoev, Nikolay and Kara, Semih and Donaghy, John and Nies, Christopher and Ersoy, O\u{g}uzhan},
  journal = {arXiv preprint arXiv:2506.10911},
  year = {2025},
}

EuroS&P

Watermarking Graph Neural Networks Based on Backdoor Attacks

Jing Xu, Stefanos Koffas, Oğuzhan Ersoy, and Stjepan Picek

In IEEE European Symposium on Security and Privacy (EuroS&P) 2023

Machine Learning Security and Privacy

Bib DOI

@inproceedings{xu2023watermarking,
  title = {Watermarking Graph Neural Networks Based on Backdoor Attacks},
  author = {Xu, Jing and Koffas, Stefanos and Ersoy, O\u{g}uzhan and Picek, Stjepan},
  booktitle = {{IEEE} European Symposium on Security and Privacy (EuroS{\&}P)},
  pages = {1179--1197},
  publisher = {{IEEE}},
  year = {2023},
}

AsiaCCS

A New Look at Blockchain Leader Election: Simple, Efficient, Sustainable and Post-Quantum

Muhammed F. Esgin, Oğuzhan Ersoy, Veronika Kuchta, Julian Loss, Amin Sakzad, Ron Steinfeld, Xiangyu Yang, and Raymond K. Zhao

In ACM Asia Conference on Computer and Communications Security (AsiaCCS) 2023

Blockchain Security and Privacy

Bib DOI

@inproceedings{esgin2023leader,
  title = {A New Look at Blockchain Leader Election: Simple, Efficient, Sustainable and Post-Quantum},
  author = {Esgin, Muhammed F. and Ersoy, O\u{g}uzhan and Kuchta, Veronika and Loss, Julian and Sakzad, Amin and Steinfeld, Ron and Yang, Xiangyu and Zhao, Raymond K.},
  booktitle = {{ACM} Asia Conference on Computer and Communications Security ({AsiaCCS})},
  publisher = {{ACM}},
  year = {2023},
}

S&P

BlindHub: Bitcoin-Compatible Privacy-Preserving Payment Channel Hubs Supporting Variable Amounts

Xianrui Qin, Shimin Pan, Arash Mirzaei, Zhimei Sui, Oğuzhan Ersoy, Amin Sakzad, Jiangshan Yu, Muhammed F. Esgin, Joseph K. Liu, and Tsz Hon Yuen

In IEEE Symposium on Security and Privacy 2023

Blockchain Security and Privacy

Bib DOI

@inproceedings{qin2023blindhub,
  title = {BlindHub: Bitcoin-Compatible Privacy-Preserving Payment Channel Hubs Supporting Variable Amounts},
  author = {Qin, Xianrui and Pan, Shimin and Mirzaei, Arash and Sui, Zhimei and Ersoy, O\u{g}uzhan and Sakzad, Amin and Yu, Jiangshan and Esgin, Muhammed F. and Liu, Joseph K. and Yuen, Tsz Hon},
  booktitle = {{IEEE} Symposium on Security and Privacy},
  pages = {},
  year = {2023},
}

Get Me out of This Payment! Bailout: An HTLC Re-routing Protocol

Oğuzhan Ersoy, Pedro Moreno-Sanchez, and Stefanie Roos

In Financial Cryptography and Data Security (FC) 2023

Blockchain Security and Privacy

Bib DOI

@inproceedings{ersoy2023bailout,
  title = {Get Me out of This Payment! Bailout: An {HTLC} Re-routing Protocol},
  author = {Ersoy, O\u{g}uzhan and Moreno-Sanchez, Pedro and Roos, Stefanie},
  booktitle = {Financial Cryptography and Data Security ({FC})},
  pages = {92--109},
  publisher = {Springer},
  year = {2023},
}

SaTML

Sniper Backdoor: Single Client Targeted Backdoor Attack in Federated Learning

Gorka Abad, Servio Paguada, Oğuzhan Ersoy, Stjepan Picek, Vı́ctor Julio Ramı́rez-Durán, and Aitor Urbieta

In IEEE Conference on Secure and Trustworthy Machine Learning (SaTML) 2023

Machine Learning Security and Privacy Decentralisation

Bib DOI

@inproceedings{abad2023sniper,
  title = {Sniper Backdoor: Single Client Targeted Backdoor Attack in Federated Learning},
  author = {Abad, Gorka and Paguada, Servio and Ersoy, O\u{g}uzhan and Picek, Stjepan and Ram{\'\i}rez-Dur{\'a}n, V{\'\i}ctor Julio and Urbieta, Aitor},
  booktitle = {{IEEE} Conference on Secure and Trustworthy Machine Learning ({SaTML})},
  pages = {},
  year = {2023},
}

AFT

SyncPCN/PSyncPCN: Payment Channel Networks without Blockchain Synchrony

Oğuzhan Ersoy, Jérémie Decouchant, Satwik Prabhu Kimble, and Stefanie Roos

In ACM Advances in Financial Technologies (AFT) 2022

Blockchain Security and Privacy

Bib DOI

@inproceedings{ersoy2022syncpcn,
  title = {SyncPCN/PSyncPCN: Payment Channel Networks without Blockchain Synchrony},
  author = {Ersoy, O\u{g}uzhan and Decouchant, J{\'e}r{\'e}mie and Kimble, Satwik Prabhu and Roos, Stefanie},
  booktitle = {{ACM} Advances in Financial Technologies ({AFT})},
  pages = {},
  year = {2022},
}

Asiacrypt

Generalized Channels from Limited Blockchain Scripts and Adaptor Signatures

Lukas Aumayr, Oğuzhan Ersoy, Andreas Erwig, Sebastian Faust, Kristina Hostáková, Matteo Maffei, Pedro Moreno-Sanchez, and Siavash Riahi

In ASIACRYPT (2) 2021

Blockchain Security and Privacy

Bib DOI

@inproceedings{DBLP:journals/iacr/AumayrEEFHMMR20,
  author = {Aumayr, Lukas and Ersoy, O\u{g}uzhan and Erwig, Andreas and Faust, Sebastian and Host{\'{a}}kov{\'{a}}, Kristina and Maffei, Matteo and Moreno{-}Sanchez, Pedro and Riahi, Siavash},
  title = {Generalized Channels from Limited Blockchain Scripts and Adaptor Signatures},
  booktitle = {{ASIACRYPT} {(2)}},
  series = {Lecture Notes in Computer Science},
  volume = {13091},
  pages = {635--664},
  publisher = {Springer},
  year = {2021},
}

S&P

Bitcoin-Compatible Virtual Channels

Lukas Aumayr, Matteo Maffei, Oğuzhan Ersoy, Andreas Erwig, Sebastian Faust, Siavash Riahi, Kristina Hostáková, and Pedro Moreno-Sanchez

In IEEE Symposium on Security and Privacy 2021

Blockchain Security and Privacy

Bib DOI

@inproceedings{DBLP:journals/iacr/AumayrEEFHMMR20a,
  author = {Aumayr, Lukas and Maffei, Matteo and Ersoy, O\u{g}uzhan and Erwig, Andreas and Faust, Sebastian and Riahi, Siavash and Host{\'{a}}kov{\'{a}}, Kristina and Moreno{-}Sanchez, Pedro},
  title = {Bitcoin-Compatible Virtual Channels},
  booktitle = {{IEEE} Symposium on Security and Privacy},
  pages = {901--918},
  publisher = {{IEEE}},
  year = {2021},
}

How to Profit from Payments Channels

Oğuzhan Ersoy, Stefanie Roos, and Zekeriya Erkin

In Financial Cryptography 2020

Blockchain Incentives

Bib DOI

@inproceedings{DBLP:conf/fc/ErsoyRE20,
  author = {Ersoy, O\u{g}uzhan and Roos, Stefanie and Erkin, Zekeriya},
  title = {How to Profit from Payments Channels},
  booktitle = {Financial Cryptography},
  series = {Lecture Notes in Computer Science},
  volume = {12059},
  pages = {284--303},
  publisher = {Springer},
  year = {2020},
}

Asiacrypt

Universal Forgery and Key Recovery Attacks on ELmD Authenticated Encryption Algorithm

Asli Bay, Oğuzhan Ersoy, and Ferhat Karakoç

In ASIACRYPT (1) 2016

Cryptography

Bib DOI

@inproceedings{DBLP:conf/asiacrypt/BayEK16,
  author = {Bay, Asli and Ersoy, O\u{g}uzhan and Karako{\c{c}}, Ferhat},
  title = {Universal Forgery and Key Recovery Attacks on {ELmD} Authenticated Encryption
                 Algorithm},
  booktitle = {{ASIACRYPT} {(1)}},
  series = {Lecture Notes in Computer Science},
  volume = {10031},
  pages = {354--368},
  year = {2016},
}