research
- Dynamic Autonomy Stacks for Safer Self-Driving
Exploring a dynamic autonomous driving stack that adapts its perception and decision-making pipeline in real time to driving conditions, with the goal of making autonomous vehicles measurably safer. Coming soon...
[ coming soon... ] - SuperFedGuard: A New Axis of Defense Against Backdoors in Federated Learning→
Weight-shared (SuperNet) federated learning quietly breaks every classical poisoning defense. SuperFedGuard turns that same weight-sharing into a defense of its own, cutting backdoor attack success by ~70% with no loss in benign accuracy.
- Finding FAULTs in Architectural Backdoors: Why Trigger Detection Fails Under Real-World Conditions→
Architectural backdoors hide malicious logic in a model's code rather than its weights — but do they actually survive deployment? FAULT shows the attacker's real-world ceiling is far lower than reported: roughly 10–20% attack success, paid for with visible accuracy damage.
- Forensic Detection and Mitigation of Architectural Backdoors in Deep Learning Models→
Framework for detecting and removing architectural backdoors (ABs) in deep learning models—malicious logic in model code rather than weights—without requiring source code. Official publication coming soon.
- Using Lightweight CNNs (U-Net Arch) for Brain Tumor Detection→
Lightweight CNN-based system using U-Net-inspired architecture for plug-and-play brain tumor detection from MRI scans, designed for hospital deployment.