CURLING

Co-Utile decentralized computing for ethics by design

A major hurdle for decentralization to become widespread is how to guarantee that all agents involved perform as expected. Whereas the risk of accidental deviation can be controlled by increasing the reliability and the redundancy of the computing and communications equipment, the most serious risk is intentional deviation. Peers that intentionally deviate may do so to attack the system or just to take advantage of it without contributing. Redundancy of agents in computations can be of use against intentional deviation, but it entails an overhead that should be minimized. It would be better if agents could be incentivized to behave as they should, which can be construed as behaving ethically.

Goals

The overarching aim of CURLING is to push forward ethics by design in decentralized machine learning in such a way that agent honesty and thus computation correctness are self-enforcing.

Ethics-by-design decentralized computing in the rational model via co-utility

This involves creating:

  • A systematic way to analyze the synergies and the conflicts between ethical values in a certain computing setting.
  • Artificial incentives for ethical behavior when such behavior is not naturally co-utile. Reputation appears as a natural way to implement artificial incentives, but other alternatives will also be explored.

Co-utile decentralized computing primitives

The focus will be on two generic primitives in the rational model:

  • Co-utile computation outsourcing, allowing a requester to delegate any computing task to untrusted but mostly rational workers with minimum redundancy.
  • Co-utile anonymous communication, allowing a sender to build an anonymous channel up to a receiver based on a variable number of hops on untrusted but mostly rational workers.

Co-utile privacy-preserving and secure federated learning

Our goal here will be to use co-utility to solve the above-mentioned conflict between accuracy, security and privacy in federated learning. Based on the co-utile anonymous communication primitive, we plan to deliver individual exact updates to the model manager, while making updates unlinkable to the workers having produced them. Furthermore, based on the co-utile outsourcing primitive, we hope to incentivize workers’ updates to be good ones, which will allow reducing worker redundancy and the detection overhead of bad updates by the model manager. Our vision is that, in this way, co-utility can help achieve both security and privacy, without losing accuracy or incurring huge overhead.

Co-utile fully decentralized learning

This objective is analogous to the previous one but for fully decentralized learning. That is, to design co-utile protocols that ensure that accuracy, security and privacy are made compatible in a learning scenario where each peer is its own model manager and uses other peers as workers that compute model updates based on their respective private data. Note that, unlike in multiparty computation, in O4 not all agents work to get their own output and hence they need other incentives. Also, unlike in computation outsourcing, here the distinction between requesters and workers is less clear.

Team

Main researchers

Dr. Josep Domingo-Ferrer

Universitat Rovira i Virgili josep.domingo (at) urv.cat 0000-0001-7213-4962

Dr. David Sánchez

Universitat Rovira i Virgili david.sanchez (at) urv.cat 0000-0001-7275-7887

Research team

  • Dr. Jordi Barrat Esteve
  • Dr. Alberto Blanco Justicia
  • Dr. Jesús A. Manjón Paniagua
  • Dr. Josep Mª Mateo Sanz

Work team

  • Dr. Markus Christen, University of Zürich, Switzerland
  • Dra. Eleonora Viganò, University of Zürich, Switzerland
  • Dr. Sergio Martínez Lluís, URV
  • Dr. Rami Haffar, URV
  • Dr. Najeeb Jebreel, URV
  • Faisal Ahmed
  • Younas Khan
  • Benet Manzanares
  • Tamim Al Mahmud

Publications

Indexed journals

  1. T. Al Mahmud, N. Jebreel, J. Domingo-Ferrer and D. Sánchez, "DP2Unlearning: An efficient and guaranteed unlearning framework for LLMs", Neural Networks, Vol. 192, 107879, Dec 2025, ISSN: 0893-6080. [DOI]
  2. R. Haffar, D. Sánchez, Y. Khan and J. Domingo-Ferrer, "Explainability-Driven Incremental Image Anonymization", Transactions on Data Privacy, Vol. 18, no. 3, pp. 135-55, Sep 2025, ISSN: 1888-5063. [PDF]
  3. R. Haffar, D. Sánchez and J. Domingo-Ferrer, "Multi-task (MTF) data set: a legally and ethically compliant collection of face images for various classification tasks", IEEE Access, Vol. 13, pp. 63827-63840, May 2025, ISSN: 2169-3536. [DOI]
  4. A. Blanco-Justicia, N. Jebreel, B. Manzanares-Salor, D. Sánchez, J. Domingo-Ferrer, G. Collell and K. E. Tan, "Digital forgetting in large language models: a survey of unlearning models", Artificial Intelligence Review, Vol. 58, 90, Apr 2025, ISSN: 0269-2821. [DOI]
  5. B. Manzanares-Salor and D. Sánchez, "Enhancing text anonymization via re-identification risk-based explainability", Knowledge-Based Systems, Vol. 310, 112945, Feb 2025, ISSN: 0950-7051. [DOI]
  6. F. Ahmed, D. Sánchez, Z. Haddi and J. Domingo-Ferrer, "MemberShield: a framework for federated learning with membership privacy", Neural Networks, Vol. 181, 106768, Jan 2025, ISSN: 0893-6080. [DOI]
  7. A. Blanco-Justicia, J. Domingo-Ferrer, N. M. Jebreel, B. Manzanares-Salor and D. Sánchez, "Unlearning in Large Language Models: We Are Not There Yet", Computer, Vol. 58, no. 1, pp. 97-100, Jan 2025, ISSN: 0018-9162. [DOI]
  8. Y. Khan, D. Sánchez and J. Domingo-Ferrer, "Federated learning-based natural language processing: a systematic literature review", Artificial Intelligence Review, Vol. 57, 320, Oct 2024, ISSN: 0269-2821. [DOI]
  9. B. Manzanares-Salor, D. Sánchez and P. Lison, "Evaluating the disclosure risk of anonymized documents via a machine learning-based re-identification attack", Data Mining and Knowledge Discovery, Vol. 38, pp. 4040-4075, Sep 2024, ISSN: 1384-5810. [DOI]
  10. L. Zhong, L. Wang, L. Zhang, J. Domingo-Ferrer, L. Xu, C. Wu and R. Zhang, "Dual-server based lightweight privacy-preserving federated learning", IEEE Transactions on Network and Service Management. Vol. 21, no. 4, pp. 1787-4800, Aug 2024, ISSN: 1932-4537. [DOI]
  11. N. Jebreel, J. Domingo-Ferrer, A. Blanco-Justicia and D. Sánchez, "Enhanced Security and Privacy via Fragmented Federated Learning", IEEE Transactions on Neural Networks and Learning Systems, Vol. 35, no. 5, pp. 6703-6717, May 2024, ISSN: 2162-237X. [DOI]
  12. N. Jebreel, J. Domingo-Ferrer, D. Sánchez and A. Blanco-Justicia, "LFighter: Defending against the label-flipping attack in federated learning", Neural Networks, Vol. 174, pp. 111-126, Feb 2024, ISSN: 0893-6080. [DOI]
  13. A. Blanco-Justicia, D. Sánchez, J. Domingo-Ferrer and K. Muralidhar, "A critical review on the use (and misuse) of differential privacy in machine learning", ACM Computing Surveys, Vol. 55, no. 8, pp. 1-16, Aug 2023, ISSN: 0360-0300. [DOI]
  14. J. Manjón, J. Domingo-Ferrer, D. Sánchez and A. Blanco-Justicia, "Secure, Accurate and Privacy-Aware Fully Decentralized Learning via Co-Utility", Computer Communications, Vol. 207, pp. 1-18, Jul 2023, ISSN: 0140-3664. [DOI]
  15. D. Sánchez, J. Domingo-Ferrer and K. Muralidhar, "Confidence-ranked reconstruction of census records from aggregate statistics fails to capture privacy risks and re-identifiability", Proceedings of the National Academy of Sciences of the United States of America, Vol. 120, no. 18, Apr 2023, ISSN: 0027-8424. [DOI]
  16. N. Jebreel and J. Domingo-Ferrer, "FL-Defender: combating targeted attacks in federated learning", Knowledge-Based Systems, Vol. 260, 110178, Apr 2023, ISSN: 0950-7051. [DOI]
  17. A. Khandpur Singh, A. Blanco-Justicia and J. Domingo-Ferrer, "Fair detection of poisoning attacks in federated learning on non-i.i.d. data", Data Mining and Knowledge Discovery, Vol. 37, no. 5, pp. 1998-2023, Mar 2023, ISSN: 1384-5810. [DOI]

Conference papers

  1. N. Jebreel, J. Domingo-Ferrer, D. Sánchez, A. Blanco-Justicia, "LFighter: Defending Against the Label-Flipping Attack in Federated Learning (Extended Abstract)", European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases 2025 - ECMLPKDD 2025, Oporto (Portugal). Core Ranking: A.
  2. F. Ahmed, D. Sánchez, J. Domingo-Ferrer and Z. Haddi, "FedBC: Privacy-preserving breast cancer diagnosis from ultrasound images using federated learning", The 22st International Conference on Modeling Decisions for Artificial Intelligence - MDAI2025, Valencia (Spain). Sep 2025. Core Ranking: B.
  3. J. Domingo-Ferrer, N. Jebreel and D. Sánchez, "Defenses against membership inference attacks on unlearned data", The 22st International Conference on Modeling Decisions for Artificial Intelligence - MDAI2025, Valencia (Spain), In Lecture Notes in Artificial Intelligence, vol. 15957, pp. 144-159, ISBN: 978-3-032-00890-9, Sep 2025. Core Ranking: B. [DOI]
  4. N. Jebreel, J. Domingo-Ferrer, A. Blanco-Justicia and D. Sánchez, "Enhanced Security and Privacy via Fragmented Federated Learning", XVIII Reunión Española sobre Criptología y Seguridad de la Información - RECSI 2024, León, Spain, Oct 2024. [PDF]
  5. R. Haffar, F. Naretto, D. Sánchez, A. Monreale and J. Domingo-Ferrer, "GLOR-FLEX: Local to Global Rule-based EXplanations for Federated Learning", IEEE World Congress on Computational Intelligence - IEEE WCCI 2024, Yokohama, Japan.Core Ranking: A. To Appear.
  6. N. Jebreel, J. Domingo-Ferrer, Y. Li, "Defending Against Backdoor Attacks by Layer-wise Feature Analysis (extended abstract)", International Joint Conference on Artificial Intelligence - IJCAI 2024, Jeju, South Korea. Core Ranking: A*. To Appear.
  7. N. Jebreel, J. Domingo-Ferrer and Y. Li, "Defending Against Backdoor Attacks by Layer-wise Feature Analysis", 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining - PAKDD2023, Osaka, Japan, In Lecture Notes in Computer Science vol. 13936, pp. 428-440, ISBN: 978-3-031-33376-7, May 2023. Best paper award. [DOI]
  8. J. Manjón and J. Domingo-Ferrer, "Computación segura multiparte coútil para cálculo de funciones arbitrarias", XVII Reunión Española sobre Criptología y Seguridad de la Información - RECSI 2022, Santander, Spain, Oct 2022. [PDF]