My research sits where machine learning meets security: how adversarial inputs break vision models, and how to defend them cheaply enough to run on the edge — work I'm now extending towards LLM systems.
LightDefend
A lightweight hybrid adversarial defense for real-time edge object detection
Object detectors deployed at the edge are exposed to adversarial inputs but cannot afford heavyweight defenses. LightDefend combines Gradient Masking with Runtime Feature Sanitization to restore detection confidence under attack while keeping latency overhead minimal — security that fits in a real-time budget.
- Threat model
- Fast Gradient Sign Method (FGSM) perturbations on input images
- Model
- YOLOv8 object detector
- Data
- COCO 2017, evaluated across 30 batches
- Defense
- Gradient Masking + Runtime Feature Sanitization
Mean detection confidence under attack, FGSM on COCO 2017 (30 batches).
Attack, then absorb.
Two lightweight defensive layers work around the detector, designed for minimal latency overhead — security that fits a real-time budget.
What an attack looks like.
Drag the budget. The perturbation stays small; its effect on a detector does not.
Interactive illustration of how FGSM builds an adversarial input — not live model output.
One gradient step, pushed in the direction that hurts the model most. Changes that look like noise to a person can collapse a detector's confidence. LightDefend is built to absorb that push at inference time, cheaply enough for the edge.
Where I'm looking.
Adversarial Machine Learning
How small, deliberate perturbations break vision models — and how to make them recover.
Edge AI Security
Defenses light enough to run in real time on constrained hardware, not just in the lab.
LLM Systems
Building and securing language-model systems: evaluation, guardrails and deployment.
Secure DevOps
Security testing as part of the pipeline — VAPT, OWASP and threat modelling by default.
On record.
Papers and preprints will appear here as they're published.
Research grounded in real attacks.
Before the thesis, industry work on the offensive and defensive side shaped how I think about threat models.