Research Monograph

Securing the future of autonomous clinical intelligence.

Technical essays on adversarial robustness in medical imaging, federated learning, and kernel-level security for clinical AI pipelines.

Research Artifacts

View all →
·Medical AI

Lipschitz-Swin3D: Adversarially Robust Volumetric MRI Segmentation

Spectrally normalized 3D Swin-UNETR architecture with certified randomized smoothing for brain tumor (BraTS) and thoracic CT organ segmentation under PGD attacks.

0.831 Robust WT Dice @ eps=2/255 (BraTS 2026)
PyTorchMONAICUDA3D ViT
·Cybersecurity

Aegis-BPF: eBPF LSM Runtime Sandbox for AI Inference & DICOM Gateways

Zero-trust Linux kernel containment daemon using eBPF LSM hooks, ring-buffer forensic telemetry, and Rust Landlock to isolate untrusted AI agents and medical imaging parsers.

<0.4% Syscall Overhead | Blocks Post-Init RCE & Egress
RusteBPF / AyaLinux LSMLandlock
·Privacy & Systems

FedMed-DP: Rényi Differentially Private Federated Hospital Training

High-performance library for multi-center clinical federated learning with per-sample gradient clipping, Rényi DP accounting, and cosine gradient-inversion auditing.

(eps = 1.8, delta = 1e-5)-DP across 12 Clinical Nodes
PythonOpacusgRPC / mTLSDifferential Privacy