Peter Lorenz

Research Scientist | Applied Machine Learning | GenAI | Agents | Ph.D. in Computer Science

Switzerland

About

Optimize and improve ML algorithms. I am currently a Postdoctoral Research Scientist at Nanyang Technological University (NTU) Singapore, focusing on trustworthy AI. I recently completed my Ph.D. from Heidelberg University, Germany, specializing in adversarial machine learning for classification, generative diffusion models with a focus on harmful outlier detection to increase the trustworthiness of AI models and prepare them for open-world problems. During my Ph.D., I conducted an internship at MIT-IBM Watson AI Lab and achieved notable results, including a top 3% paper award at the ICASSP conference, a top-20 ranking in the CVPR 2022 Art-of-Robustness Challenge, and two times Oxford summer school acceptance. In addition to my research, I actively contribute to the machine learning community as a reviewer for top-tier conference workshops such as ICML, ICLR, and NeurIPS, but also ICASSP (ML track) and EuroCrypt (ML track). My passion lies in advancing the field of machine learning, with a particular focus on its practical applications. Previously, I worked in academia and industry on the CARLA simulator for pedestrian safety and classification models for mobile robotics, including vehicles and drones. Before that, I contributed to the vision system of autonomous robots (Team TEDUSAR @TU Graz, which won the RoboCup competition "best in autonomy" in 2016.) in terms of my Bachelor thesis. In future, I would be interested in adaptive and evolving AI, in particularly in industry setting such as tech or finance.

Experience

  • Applied Scientist at Idiap Research Institute
    Feb 2026 - Present · 5 mos

    Foundation model fine-tuning for static and time-series data for PAD (presentation attack detection) - face spoofing detection. AI coding with cursor (claude, opus, sonnet, codex, ...).

  • Research Scientist at Nanyang Technological University Singapore
    Aug 2024 - Aug 2025 · 1 yr 1 mo

    Improved Google-Research's model theft [1, 2] approach [3]—more accurate and deeper layer reconstruction of model weights. The research community assumed that this was not the case [4]. We proved otherwise. Resulted in a EUROCRYPT publication. I am fortunate to work with Prof. Thomas Peyrin. Ranked as the 2nd best in AI worldwide (usnews.com), 30k students and staff. [1] https://arxiv.org/pdf/2506.17047 [2] https://owasp.org/www-project-machine-learning-security-top-10/docs/ML05_2023-Model_Theft [3] https://github.com/google-research/cryptanalytic-model-extraction [4] https://www.youtube.com/watch?v=PfoYLmbyOQE - Stealing Weights of a Production LLM Like OpenAI’s ChatGPT with Nicholas Carlini

  • Research Associate at Fraunhofer ITWM
    Feb 2021 - Jul 2024 · 3 yrs 6 mos

    Robust Computer Vision. Pattern Recognition and Data Science. GenAI such as Diffusion Models. Multimodality (text-to-image). - I showed with empirical evaluations that AutoAttack's perturbation is not the best choice for an adversarial attack. - Created as first-time prompts to neutralize harmful, manipulated data. - Created a diffusion model deepfake dataset. - I found that the latent spaces of diffusion models are misaligned with the learned manifold. Open-source code: - SpectralDefense: https://github.com/adverML/SpectralDef_Framework - OpenOOD: https://github.com/Jingkang50/OpenOOD/pull/275 I am grateful to my advisor, Prof. Janis Keuper, and my team lead.

  • AI Research Intern - MIT-IBM Watson AI Lab - Visiting Research at IBM
    Jul 2022 - Sep 2022 · 3 mos

    Published the paper "Visual prompting for adversarial robustness" at NeurIPS workshops and ICASSP where it got recognition in the top 3%. The idea of visual prompting is derived from Large Language Model (LLM) prompting. It is the first attempt with test-time prompt selection. https://research.ibm.com/publications/visual-prompting-for-adversarial-robustness I am very grateful for the supervision of Prof. Sijia Liu and Dr. Pin-Yu Chen.

  • Research And Development Engineer at MBDA
    Aug 2020 - Jan 2021 · 6 mos

    Top 1% Leading Employer Germany Data cleaning and preprocessing on 1000 samples for initial successful training. PoC by fine-tuning deep learning (time series, LSTM) vision models (Yolo, MobileNet) for drone applications for the FCAS project (Taurus) within a few months.