Najeeb Jebreel
Juan de la Cierva Postdoctoral Fellow, KISON research group, Universitat Oberta de Catalunya (UOC)
I study what trained machine learning models reveal about the data they were trained on, and how to measure, limit and remove that leakage.
Google Scholar ORCID GitHub LinkedIn ResearchGate njebreel@uoc.edu
News
A Critical Review on the Effectiveness and Privacy Threats of Membership Inference Attacks is published in the ESORICS 2026 proceedings (Rome, CORE A).
Co-authored paper Comparative Heatmap Explanations for Deep Learning-Based Breast Density Classification published at IEEE eSmarTA 2026.
Revisiting the LiRA Membership Inference Attack Under Realistic Assumptions is published in PoPETs 2026, Issue 3, with all three artifact badges: Available, Functional and Reproduced.
Started a Juan de la Cierva Postdoctoral Fellowship at the Universitat Oberta de Catalunya (UOC).
Membership Inference Attacks Beyond Overfitting, presented at DPM 2025, is published in the ESORICS 2025 workshop proceedings (LNCS).
Joined the program committee of PoPETs 2027.
Earlier news
Survey Digital Forgetting in Large Language Models published in Artificial Intelligence Review (Q1, first decile).
Presented the PAKDD 2023 best paper at IJCAI 2024 (CORE A*), Sister Conferences track, Jeju, South Korea.
Enhanced Security and Privacy via Fragmented Federated Learning published in IEEE TNNLS (Q1, first decile).
Received the URV Extraordinary Doctorate Award for the PhD thesis in Computer Engineering and Mathematics of Security.
Best Paper Award at PAKDD 2023 (CORE A), Osaka, from 869 submissions.
About
Najeeb Jebreel is a Juan de la Cierva Postdoctoral Fellow in the KISON research group at the Universitat Oberta de Catalunya (UOC). His research is on privacy and security in machine learning: membership inference and privacy auditing, machine unlearning, differential privacy, and attacks and defences in federated learning.
Before joining UOC in July 2026, he was a postdoctoral researcher in the CRISES research group at Universitat Rovira i Virgili (URV), where he received his PhD in 2023 cum laude with the URV Extraordinary Doctorate Award. He is affiliated with the UNESCO Chair in Data Privacy at URV.
His work has appeared in PoPETs, ESORICS, IEEE TNNLS, Neural Networks, Knowledge-Based Systems and Artificial Intelligence Review, and has been cited more than 720 times (h-index 11, Google Scholar). His paper on layer-wise feature analysis against backdoor attacks received the Best Paper Award at PAKDD 2023 and was invited to IJCAI 2024.
He serves on the program committee of PoPETs 2027, has supervised three Master's theses and currently supervises two PhD students. He has collaborated with Huawei Technologies and the Notre Dame–IBM Technology Ethics Lab, and has contributed to projects funded by the European Commission, INCIBE and the Spanish State Research Agency.
Research
-
Privacy auditing and membership inference
Membership inference attacks are the standard tool for measuring what a model leaks about its training data. I re-evaluate them under realistic conditions: models trained with ordinary anti-overfitting practice, decision thresholds calibrated without access to the target's data, skewed membership priors, and reproducibility across training runs. Under these conditions even the strongest attacks, such as LiRA, are much weaker and less reliable than usually reported. I also study why a small set of samples stays vulnerable in models that generalise well.
Key papers: LiRA (PoPETs 2026), Critical review (ESORICS 2026), Beyond overfitting (DPM 2025)
-
Machine unlearning
Data protection law gives people the right to have their data forgotten, including by models trained on it. I develop unlearning methods with formal guarantees: models are first trained on data protected with differential privacy or k-anonymity, so later forgetting requests can be served cheaply with guarantees inherited from the privacy model. I also study unlearning in large language models and how to check whether forgetting actually happened.
Key papers: Survey (AI Review 2025), DP2Unlearning, EUPG, MIA defenses for unlearned data (MDAI 2025)
-
Security of federated and centralised learning
Federated learning trains models without pooling raw data, but it opens new attack surfaces. I have built defences against label-flipping, Byzantine and backdoor attacks, proposed fragmented federated learning to improve privacy and security together, and introduced a layer-wise feature analysis for detecting backdoors that received the PAKDD 2023 Best Paper Award. Earlier work covers watermarking deep learning models to prove ownership.
Key papers: LFighter, FL-Defender, Fragmented FL (IEEE TNNLS), Layer-wise feature analysis (PAKDD 2023), KeyNet
Publications
2026
-
Revisiting the LiRA Membership Inference Attack Under Realistic Assumptions
Proceedings on Privacy Enhancing Technologies (PoPETs), 2026(3), pp. 765–786
Artifact badges: Available, Functional and Reproduced.
-
A Critical Review on the Effectiveness and Privacy Threats of Membership Inference Attacks
Computer Security – ESORICS 2026, Rome, LNCS 16969, pp. 45–64, Springer
-
Comparative Heatmap Explanations for Deep Learning-Based Breast Density Classification
6th International Conference on Emerging Smart Technologies and Applications (eSmarTA 2026), IEEE
-
Membership Inference Attacks Beyond Overfitting
Computer Security. ESORICS 2025 International Workshops (DPM 2025), Toulouse, LNCS 16231, pp. 32–48, Springer
2025
-
DP2Unlearning: An Efficient and Guaranteed Unlearning Framework for LLMs
Neural Networks, vol. 192, art. 107879, Elsevier
-
Digital Forgetting in Large Language Models: A Survey of Unlearning Methods
Artificial Intelligence Review, vol. 58, no. 3, art. 90, Springer
-
Unlearning in Large Language Models: We Are Not There Yet
Computer, vol. 58, no. 1, pp. 97–100, IEEE
-
Defenses Against Membership Inference Attacks on Unlearned Data
Modeling Decisions for Artificial Intelligence (MDAI 2025), València, LNCS 15957, pp. 145–159, Springer
-
Efficient Unlearning with Privacy Guarantees
arXiv preprint arXiv:2507.04771
2024
-
LFighter: Defending against the Label-Flipping Attack in Federated Learning
Neural Networks, vol. 170, pp. 111–126, Elsevier
-
Enhanced Security and Privacy via Fragmented Federated Learning
IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 5, pp. 6703–6717
-
Defending Against Backdoor Attacks by Layer-wise Feature Analysis (Extended Abstract)
Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI 2024), Sister Conferences Best Papers track, pp. 8416–8420, Jeju
-
An Examination of the Alleged Privacy Threats of Confidence-Ranked Reconstruction of Census Microdata
Privacy in Statistical Databases (PSD 2024), pp. 213–224, Springer
2023
-
FL-Defender: Combating Targeted Attacks in Federated Learning
Knowledge-Based Systems, vol. 260, art. 110178, Elsevier
-
Defending Against Backdoor Attacks by Layer-wise Feature Analysis
27th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD 2023), Osaka, pp. 428–440, Springer
869 submissions, 16.5% acceptance rate. Invited to the IJCAI 2024 Sister Conferences track.
-
Protecting Models and Data in Federated and Centralized Learning
PhD thesis, Universitat Rovira i Virgili, supervised by Josep Domingo-Ferrer
2022
-
Generating Deep Learning Model-Specific Explanations at the End User's Side
International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 30, Supp02, pp. 255–278
-
Generation of Synthetic Trajectory Microdata from Language Models
Privacy in Statistical Databases (PSD 2022), pp. 172–187, Springer
-
Measuring Fairness in Machine Learning Models via Counterfactual Examples
Modeling Decisions for Artificial Intelligence (MDAI 2022), pp. 119–131, Springer
2021
-
KeyNet: An Asymmetric Key-Style Framework for Watermarking Deep Learning Models
Applied Sciences, vol. 11, no. 3, art. 999, MDPI
-
Explaining Image Misclassification in Deep Learning via Adversarial Examples
Modeling Decisions for Artificial Intelligence (MDAI 2021), pp. 323–334, Springer
2020
-
Co-Utile Peer-to-Peer Decentralized Computing
20th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid 2020), pp. 31–40, IEEE
-
Efficient Detection of Byzantine Attacks in Federated Learning Using Last Layer Biases
Modeling Decisions for Artificial Intelligence (MDAI 2020), pp. 154–165, Springer
-
Detecting Bad Answers in Survey Data Through Unsupervised Machine Learning
Privacy in Statistical Databases (PSD 2020), pp. 309–320, Springer
Citation counts from Google Scholar, November 2026.
Projects
Current
- 2026 –
Juan de la Cierva Postdoctoral Fellowship
MCIU/AEI and FSE+, Spanish Government
Research on membership inference, machine unlearning and differential privacy at UOC.
Earlier
- 2024
Mitigating Ethical Risks in LLMs through Localized Unlearning
Notre Dame–IBM Technology Ethics Lab
Selective forgetting in LLMs to protect privacy and reduce harmful content generation. - 2023 – 2025
Strategic Project HERMES
INCIBE and URV
Advanced cybersecurity and privacy-preserving technologies. - 2022 – 2025
CURLING: Co-Utile decentRaLized computING
Spanish Ministry of Science and Innovation
Ethics by design for distributed, privacy-first computing systems. - 2021 – 2024
CRISES: Security and Privacy (2021 SGR 00115)
AGAUR, Generalitat de Catalunya
Researcher in the CRISES security and privacy group at URV. - 2021 – 2024
MobiDataLab
European Commission, Horizon 2020
Privacy-preserving techniques for sharing mobility data. - 2020 – 2023
SoBigData++
European Commission, Horizon 2020
Privacy and security for the European big data analytics infrastructure.
CV
Experience
- 2026 – now
Juan de la Cierva Postdoctoral Fellow
Universitat Oberta de Catalunya (UOC)
Membership inference, machine unlearning and differential privacy. - 2024 – 2026
Postdoctoral Researcher
Universitat Rovira i Virgili (URV)
ML security, privacy-preserving AI and LLM unlearning. - 2023
Research Support Technician
Universitat Rovira i Virgili (URV)
Technical support for cybersecurity and privacy projects. - 2020 – 2023
FI Predoctoral Researcher
Universitat Rovira i Virgili (URV)
Doctoral research on ML security and privacy (grant FI_B00760). - 2012 – 2017
Computer Systems Manager
Hodeidah University, Yemen
University IT systems and infrastructure.
Education
- 2020 – 2023
PhD in Computer Engineering and Mathematics of Security
Universitat Rovira i Virgili, Spain
Cum laude, URV Extraordinary Doctorate Award - 2019
MSc in Computer Security Engineering and Artificial Intelligence
Universitat Rovira i Virgili, Spain
GPA 8.1 / 10 - 2009
BSc in Computer Science and Engineering
Hodeidah University, Yemen
GPA 79.39 / 100
Awards
- 2025
Juan de la Cierva 2024 Postdoctoral Fellowship
Spanish Government. Awarded October 2025, started July 2026 at UOC. - 2024
Invited talk, IJCAI 2024 Sister Conferences track
Based on the PAKDD 2023 best paper (IJCAI is CORE A*). - 2023
Best Paper Award, PAKDD 2023
27th Pacific-Asia Conference on Knowledge Discovery and Data Mining (CORE A); 869 submissions, 16.5% acceptance rate. - 2023
URV Extraordinary Doctorate Award
PhD in Computer Engineering and Mathematics of Security.
Supervision
- 2023 –
PhD students
Supervising 2 PhD students (ongoing). Master's theses
Supervised 3 Master's theses, all defended.
Service
- 2026 – now
Program committee member, PoPETs 2027
Privacy Enhancing Technologies Symposium (CORE A). - 2020 – now
Journal reviewer
ACM Computing Surveys, ACM TIST, IEEE TNNLS, IEEE TIFS, IEEE TDSC, IEEE Internet of Things Journal, Artificial Intelligence Review, Knowledge-Based Systems, Future Generation Computer Systems, Applied Sciences and Electronics. - 2020 – now
Conference reviewer
ACM CCS, WPES 2020 (co-located with CCS), ESORICS, IEEE ICDM and ACISP. - 2020 – now
Research collaborations
Huawei Technologies and the Notre Dame–IBM Technology Ethics Lab, on backdoor defences, LLM unlearning and privacy-preserving ML.
Contact
Office
KISON Research GroupUniversitat Oberta de Catalunya (UOC)
Rambla del Poblenou, 156
08018 Barcelona, Spain