Securing the Future of Autonomous Clinical Intelligence.
I am a principal researcher and security engineer working at the convergence of Medical Artificial Intelligence and Offensive/Defensive Cybersecurity.
Modern healthcare increasingly relies on deep neural architectures—from 3D Vision Transformers parsing volumetric CT/MRI scans to clinical foundation models assisting diagnostic triage. Yet these models operate inside interconnected hospital networks and remain vulnerable to imperceptible adversarial perturbations, DICOM parser memory corruption, and gradient inversion attacks.
My research focuses on three core problems: (1) provably robust 3D medical segmentation under bounded L-infinity perturbations, (2) kernel-level eBPF & Rust sandboxing for untrusted AI inference workloads, and (3) differentially private federated learning across multi-institutional clinical cohorts.
Research Focus Areas
Adversarial Robustness in 3D Volumetric Imaging
Designing certified randomized smoothing and spectral-regularized 3D Vision Transformers (Swin-UNETR) resilient to PGD and frequency-domain perturbations in MRI/CT volumes.
eBPF Runtime Telemetry & Memory-Safe Sandboxing
Eliminating parser exploits in clinical DICOM gateways and isolating untrusted model execution using Linux Landlock, seccomp-bpf, and ring-buffer eBPF probes.
Federated Clinical Learning & Gradient Defense
Preventing patient reconstruction attacks from shared model gradients using Rényi Differential Privacy (RDP) and secure aggregation protocols.
Experience
Leading cross-disciplinary research on adversarial robustness for FDA SaMD radiology pipelines and zero-trust inference enclaves.
Architected eBPF-based runtime enforcement for GPU compute clusters and discovered multiple critical CVEs in medical imaging parsers.
Dissertation on certified robustness bounds for volumetric brain tumor segmentation in multi-modal 3T MRI.
Selected Publications
Contact
For coordinated vulnerability disclosures in medical devices or clinical AI pipelines, academic collaborations, or editorial inquiries, use encrypted email below. Content on this page is managed via Payload CMS (/admin).