Adam Dziedzic

Adam Dziedzic is a Tenure Track Faculty Member at CISPA Helmholtz Center for Information Security, where he co-leads the SprintML group. His research focuses on Secure, Private, Robust, Interpretable, and Trustworthy Machine Learning, with applications to diffusion models, large language models, and collaborative learning.

He previously held a Postdoctoral Fellow position at the Vector Institute and the University of Toronto, advised by Prof. Nicolas Papernot. He earned his PhD in computer science from the University of Chicago, advised by Prof. Sanjay Krishnan, and holds BSc and MSc degrees from Warsaw University of Technology in Poland.

His work spans privacy attacks on image autoregressive models, dataset inference for large language models, differentially private collaborative learning, and defenses against model extraction. Recent publications (2025–2024) address privacy leakage in generative models, synthetic data calibration, secure noise sampling protocols, and memorization mechanisms in self-supervised and multi-modal architectures.

Key contributions include:

• Developing POST for efficient soft prompt transfer in LLMs with privacy guarantees
• Proposing NeMo to localize and mitigate memorization in diffusion models
• Advancing Dataset Inference methods for diffusion and self-supervised models
• Designing secure multiparty computation protocols for collaborative learning

His research leverages techniques like differential privacy, knowledge distillation, and statistical testing to balance utility and privacy across distributed systems.