Muhammad Huzaifa

I am a doctoral researcher at the CISPA Helmholtz Center for Information Security , where I work closely with Thorsten Eisenhofer and Lea Schönherr . I am passionate about adversarial machine learning, with the goal of developing AI systems that extend human intelligence in a secure and privacy-preserving manner. My current research focuses on the safety and security of generative AI.

Prior to my current role, I completed my Master’s degree at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) under the supervision of Prof. Fahad Khan and Prof. Salman Khan . During this time, my research spanned multimodal representation learning, the robustness of vision models, and the adaptation of vision-language models to low-data settings, including few-shot and zero-shot scenarios. Before that, I earned my undergraduate degree in Electrical Engineering from NUST, Islamabad. Along the way, I also worked as a researcher at the Tübingen AI Center and MBZUAI .

I am always open to collaborations and discussions on research, security, privacy, or related topics. Feel free to email me.

Email  /  Linkedin  /  Github /  Scholar  /  Twitter  /  Bluesky

News

We show that learned latent communication links can compromise the safety of multi-agent systems built from safety-aligned agents, and that updating the links alone can repair them.

WISE-ATTA decides online when to request labels under a fixed annotation budget, matching or outperforming recent active test-time adaptation methods with far fewer labels.

VisualOverload introduces a dense-scene benchmark for probing detailed visual understanding in vision-language models, revealing substantial performance gaps on crowded, high-detail images.

ObjectCompose evaluates how robust vision-based models are to object-to-background compositional shifts, providing a benchmark for studying context sensitivity in modern visual recognition systems.

Miscellanea