Leo Schwinn

Leo Schwinn is a Lecturer at the Technical University of Munich (TUM) within the Department of Computer Science (I26), working in the Data Analytics and Machine Learning group supervised by Prof. Stephan Günnemann at the TUM School of Computation, Information and Technology. His research focuses on robust machine learning with particular emphasis on data-efficient learning and robustness vulnerabilities of Large Language Models (LLMs).

Dr. Schwinn's research interests span multiple critical areas in contemporary machine learning including:

• Robustness against adversarial attacks in LLMs
• Embedding space vulnerabilities and defenses
• Model unlearning and privacy preservation
• Efficient training methodologies for large models
• Time-series forecasting with probabilistic frameworks
• Graph-based machine learning approaches

His work bridges theoretical understanding with practical security implications of modern AI systems.

Analysis of his recent publications (2023-2025) reveals a strong focus on LLM security, with multiple papers accepted at premier conferences including ICML, CVPR, ICLR, and NeurIPS. His research demonstrates consistent innovation in identifying novel attack vectors while developing practical defense mechanisms, particularly through embedding space manipulation techniques. The work shows increasing sophistication in handling both theoretical aspects of model robustness and practical deployment concerns.

His notable scientific achievements include:

• Receiving the ATE dissertation price for his PhD work at FAU
• Securing an oral presentation at ICLR 2025
• Organizing the ICLR BlogPost Track
• Becoming a member of ELLIS (European Laboratory for Learning and Intelligent Systems)

Dr. Schwinn has served as review process chair for the 2024 Conference on Lifelong Learning Agents (CoLLAs) and actively collaborates with researchers at Mila Quebec AI Institute. His research group at TUM focuses on addressing fundamental challenges in machine learning robustness, particularly as they apply to real-world deployment scenarios where security and reliability are paramount.

He maintains active GitHub repositories related to LLM security research, including circuit-breakers-eval and LLM_Embedding_Attack, demonstrating his commitment to open science and reproducible research in the field of AI security.