CV

Education

• Ph.D. in Computer Science
  Warsaw University of Technology • Oct 2021 – Oct 2025 (awaiting defence)
      Thesis: “Reliable and Safe Generative Models”

• M.Sc. in Computer Science
  Warsaw University of Technology • Feb 2018 – Jun 2021
      Specialization: Information and Decision Systems
      Erasmus Exchange – University of Vienna, Mar 2020 – Jun 2020

• M.Sc. in Power Engineering
  Warsaw University of Technology • Feb 2018 – Sep 2020
      Erasmus Exchange – Polytechnic University of Catalonia, Sep 2018 – Jan 2019

• B.Sc. in Quantitative Methods in Economics and Information Systems
  Warsaw School of Economics • Oct 2016 – Feb 2020

• B.Sc. in Power Engineering
  Warsaw University of Technology • Oct 2014 – Feb 2018

Work experience

• Astra Fellow, Constellation Institute at Berkeley, USA • Jan 2025 – Present
  Working with Owain Evans’s TruthfulAI on safety and alignment of Large Language Models

• MARS Fellow, Cambridge AI Safety Hub • Dec 2022 – Jan 2026
  Research on safety of Vision-Language Models with Yossi Gandelsman

• Senior Specialist, NASK National Research Institute • May 2025 – Present
  Department of Security and Transparency of Artificial Intelligence

• Visiting Researcher, CISPA – Helmholtz Center for Information Security • Mar 2024 – May 2024
  Working on identifying the use of copyrighted data in Diffusion Models with Adam Dziedzic

• PhD Student Researcher, IDEAS NCBR • Oct 2023 – Apr 2024
  Research in Systems Security and Data Privacy

• Collaborator, CERN, the European Organization for Nuclear Research • Feb 2019 – Present
  Fast simulations of a particle calorimeter at the Large Hadron Collider using generative adversarial networks and variational autoencoders

• Junior Machine Learning Engineer, Samsung R&D Institute Poland • Jul 2019 – Oct 2021
  Development of machine learning systems for automatic code repair

• IAESTE Internship, National Technical University in Bahia Blanca, Argentina • Aug 2017 – Sep 2017
  Development of grid power failure simulations

Skills

• Strong background in machine learning, with a focus on generative models and their safety, including Large Language Models, Generative Vision Models, and Vision-Language Models
• Practical research experience across academia and international institutions
• Experience in frontier model adaptation and evaluation
• Advanced programming proficiency in Python 3
• In-depth knowledge of:
  • Machine Learning: PyTorch, TensorFlow, Keras
  • Data Science: Pandas, NumPy
  • Data Visualization: Matplotlib, Seaborn
• Practical knowledge of:
  • Version control: Git
  • Operating systems: Linux
  • Containerization: Docker
• Foreign languages: English C2 (fluent), German B1, Japanese A2

Selected Publications

• Universal Properties of Activation Sparsity in Modern Large Language Models
  F. Szatkowski, P. Będkowski, A. Devoto, J. Dubiński, P. Minervini, M. Piórczyński, S. Scardapane, B. Wójcik
  International Conference on Learning Representations (ICLR), 2026, CORE A*

• On Stealing Graph Neural Network Models
  M. Podhajski, J. Dubiński, F. Boenisch, A. Dziedzic, A. Pręgowska, T. Michalak
  AAAI Conference on Artificial Intelligence (AAAI), 2026, CORE A*

• Privacy Attacks on Image Autoregressive Models
  A. Kowalczuk*, J. Dubiński*, F. Boenisch, A. Dziedzic
  *International Conference on Machine Learning (ICML), 2025, CORE A***

• CDI: Copyrighted Data Identification in Diffusion Models
  J. Dubiński*, A. Kowalczuk*, F. Boenisch, A. Dziedzic
  *Conference on Computer Vision and Pattern Recognition (CVPR), 2025, CORE A***

• Learning Graph Representation of Agent Diffuser
  Y. Djenouri, N. Belmecheri, T. Michalak, J. Dubiński, A. N. Belbachir, A. Yazidi
  International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2025, CORE A*

• Bucks for Buckets (B4B): Active Defenses Against Stealing Encoders
  J. Dubiński*, S. Pawlak*, F. Boenisch*, T. Trzcinski, A. Dziedzic
  Neural Information Processing Systems (NeurIPS), 2023, CORE A*

• Towards More Realistic Membership Inference Attacks on Large Diffusion Models
  J. Dubiński, A. Kowalczuk, S. Pawlak, P. Rokita, T. Trzciński, P. Morawiecki
  Winter Conference on Computer Vision (WACV), 2024, CORE A

• Efficient Model-Stealing Attacks Against Inductive Graph Neural Networks
  M. Podhajski, J. Dubiński, F. Boenisch, A. Dziedzic, A. Pregowska, T. Michalak
  European Conference on Artificial Intelligence (ECAI), 2024, CORE A

• Selectively Increasing the Diversity of GAN-generated Samples
  J. Dubiński, K. Deja, S. Wenzel, P. Rokita, T. Trzciński
  International Conference on Neural Information Processing (ICONIP), 2022, CORE A

• Progressive Latent Replay for Efficient Generative Rehearsal
  S. Pawlak, F. Szatkowski, M. Bortkiewicz, J. Dubiński, T. Trzciński
  International Conference on Neural Information Processing (ICONIP), 2022, CORE A

• Efficient LLM Moderation with Multi-Layer Latent Prototypes
  M. Chrabąszcz, F. Szatkowski, B. Wójcik, J. Dubiński, T. Trzciński
  ICLR Workshop on Building Trust in Language Models and Applications, 2025