Bogdan Kulynych

Email
  • bogdan.kulynychepfl.ch
    All things academic
  • hellobogdankulynych.me
    All things personal
Github bogdan-kulynych

Graduate student at EPFL Security and Privacy Engineering Lab (SPRING). Former intern at Google and CERN.

I am interested in the intersection of privacy, security, and machine learning.

Projects and publications

Protective optimization technologies

  • Protective optimization technologies (POTs)—a concept of developing technologies that counteract harmful optimization systems from outside.

    Rebekah Overdorf, Bogdan Kulynych, Ero Balsa, Carmela Troncoso, Seda Gürses. POTs: Protective Optimization Technologies. Under submission.

    See accompanying code that implements poisoning attacks and adversarial examples against a ML-based credit scoring system.

  • textfool—a tool for evading text classifiers, with the goal of countering privacy-invasive classifiers.

Provable security and privacy for machine learning

  • Provably minimal-effort evasion attacks in constrained discrete domains.

    Bogdan Kulynych, Jamie Hayes, Nikita Samarin, Carmela Troncoso. Evading classifiers in discrete domains with provable optimality guarantees. NIPS 2018 Workshop on Security in Machine Learning, December 7, 2018, Montreal, Canada.

  • mia—an open-source library for evaluating Keras or PyTorch models against attacks on privacy of the training data: membership inference attacks.

  • Power indices for machine learning models—game-theoretic tools that can be used to prove that certain features can not influence a classifier’s decision.

    Bogdan Kulynych, Carmela Troncoso. Feature importance scores and lossless feature pruning using Banzhaf power indices. NIPS 2017 Symposium on Interpretable Machine Learning, December 7, 2017, Long Beach, CA, USA.

Privacy-preserving cryptographic data structures

Academic service