At Technical University of Darmstadt - System Security Lab, I design machine-learning models for security-critical and multimodal data, and defenses against backdoors, data leakage, and adversarial behavior. I also build evaluation pipelines for LLM safety and robustness.
I designed VoiceRadar, a system for detecting AI-generated and manipulated speech. Its paper received the Distinguished Paper Award at NDSS 2025, and the system is being commercialized with an industry partner.
For Light2Lie, I developed a physics-inspired vision pipeline that uses reflectance, illumination, and signal inconsistencies to distinguish AI-generated images from real ones. I evaluated its robustness under distribution shifts and adversarial perturbations.
In my current AI Security Research Fellowship, I’m testing whether a reasoning model’s hidden context can be recovered from its output logits. I also built a LoRA fine-tuning pipeline for studying self-narrative language drift and emergent misalignment.
