Berkant Turan

ML Researcher @ IOL Lab at ZIB | PhD Candidate @ TU Berlin | Member @ Berlin Mathematical School

profile.jpg

I am a third-year PhD candidate at the Institute of Mathematics, TU Berlin and a research associate at Zuse Institute Berlin (ZIB). Under the supervision of Prof. Sebastian Pokutta, I am currently part of the Interactive Optimization and Learning (IOL) research group at ZIB. Additionally, I am a member of the Berlin Mathematical School (BMS), which is part of the Math+ Excellence Cluster.

My research interests focus on the interpretability, robustness, and safe deployment of neural networks in high-stakes applications. I develop interactive, multi-agent models that provide provable insights into the decision-making processes of black-box systems. By utilizing feature selectors within an adversarial framework, I aim to expose the reasoning behind model predictions, addressing core challenges in the interpretability and security of complex AI systems.

Additionally, I am interested in the interconnections between model security approaches, such as adversarial robustness and backdoor-based watermarks, and the theoretical limits tied to these techniques in different learning tasks. My work also investigates transferable attacks, which exploit vulnerabilities across multiple defenses using cryptographic tools, uncovering fundamental links between AI security and cryptography.

Before starting my PhD, I focused on Deep Hybrid Discriminative-Generative Modeling, investigating the optimization and behavior of Variational Autoencoders and Residual Networks for out-of-distribution detection, robustness, and calibration in computer vision tasks.

news

10/2024 Excited to announce that The Good, the Bad and the Ugly: Watermarks, Transferable Attacks and Adversarial Defenses is now available on arXiv! Many thanks to my collaborators, Grzegorz Głuch (EPFL at the time), Sai Ganesh Nagarajan (ZIB) and Sebastian Pokutta (ZIB), for their contributions to this project!
06/2024 Unified Taxonomy of AI Safety: Watermarks, Adversarial Defenses and Transferable Attacks got accepted at ICML 2024 Workshop on Theoretical Foundations of Foundation Models. See you in Vienna!
03/2024 Our recent paper, Interpretability Guarantees with Merlin-Arthur Classifiers, has been accepted at AISTATS 2024. Looking forward to meeting you in Valencia.
01/2024 I am delighted to have become a member of the Math+ Excellence Cluster working on the emerging field EF1 – Extracting dynamical Laws from Complex Data.
07/2023 Our PhD proposal, Extending Merlin-Arthur Classifiers for Improved Interpretability, won the Best Proposal Award at the 1st xAI World Conference in Lisbon, Portugal.
09/2022 Excited to have started my PhD at TU Berlin and the Zuse Institute Berlin in the Interactive Optimization and Learning research lab, under the supervision of Sebastian Pokutta.

selected publications

  1. gluch2024_goodbadugly.png
    The Good, the Bad and the Ugly: Watermarks, Transferable Attacks and Adversarial Defenses
    Grzegorz Głuch, Berkant Turan, Sai Ganesh Nagarajan, and Sebastian Pokutta
    arXiv preprint arXiv:2410.08864, 2024
  2. icml2024_poster.png
    Unified Taxonomy in AI Safety: Watermarks, Adversarial Defenses, and Transferable Attacks
    Grzegorz Gluch, Sai Ganesh Nagarajan, and Berkant Turan
    ICML 2024 Workshop on Theoretical Foundations of Foundation Models (TF2M), 2024
  3. merlin-arthur-classifier.jpeg
    Interpretability Guarantees with Merlin-Arthur Classifiers
    Stephan Wäldchen, Kartikey Sharma, Berkant Turan, Max Zimmer, and 1 more author
    In Proceedings of The 27th International Conference on Artificial Intelligence and Statistics (AISTATS), PMLR, 2024
  4. world_conference_on_explainable_artificial_intelligence_logo.jpeg
    Extending Merlin-Arthur Classifiers for Improved Interpretability
    Berkant Turan
    In Joint Proceedings of the xAI-2023 Late-breaking Work, Demos and Doctoral Consortium, co-located with the 1st World Conference on eXplainable Artificial Intelligence (xAI-2023), Jul 2023
    (Best Proposal Award)