Autonomous Vehicles and UAVs: CAN Intrusion Detection, GPS Spoofing, V2X, Drone Safety

projects

Hybrid Clustering-Transformer IDS for Rural CAVs

under submissionco-advised · Mahsa Tavasoli (lead)

An edge-first intrusion detection system for connected and automated vehicles in rural areas, where intermittent connectivity makes cloud-based defense unreliable. It runs fast per-CAN-ID unsupervised clustering for local screening, then a compact transformer for byte-sequence context, fused with confidence-weighted, per-ID adaptive thresholds so alert sensitivity adapts to each message identifier.

100% recall and 98.9% accuracy on flooding attacks, connectivity-independent, false positives concentrated on one CAN ID for targeted calibration

keywords: CAN bus, intrusion detection, connected vehicles, edge AI, rural ITS

On the Fundamental Limits of GPS Spoofing Detection

VehicleSec 2026under reviewco-advised research

A theory result for autonomous-vehicle localization security. Every position-domain spoofing detector compares the GPS fix to an IMU dead-reckoning estimate and flags disagreement. This work derives a closed-form lower bound showing reliable detection requires the accumulated spoofing displacement over a window to exceed roughly three times the baseline noise. Published 90%+ F1 numbers turn out to be artifacts of favorable test conditions.

A $223 software-defined radio steers a vehicle off-lane in 20 s; on highways with consumer GPS a gradual 0.5 m/s attack falls below the detectable limit

keywords: GPS spoofing, GNSS security, autonomous vehicle localization, detection limits

VIN2VICTIM: From a Single VIN to Targeted Tracking

USENIX VehicleSec 2026first author · with Mohammad Sepahi

A connected-vehicle privacy attack that converts a single Vehicle Identification Number into targeted physical tracking through four composable stages: owner deanonymization, trajectory re-identification, home and workplace inference, and departure prediction. It shows vehicles are structurally worse than phones for privacy, because VINs are immutable lifelong correlators, consent is coarse and vehicle-scoped, and VINs bridge the digital-to-physical gap.

Two spatio-temporal points re-identify 96 to 99% of taxi traces and 82% of personal traces; the attack survives 60-minute coarsening despite 91% data reduction

keywords: VIN, vehicle privacy, telematics, data brokers, re-identification

Data Communication Challenges of CAVs in Rural Areas

IEEE Access 2025co-advised · Mahsa Tavasoli (lead)

A published review of the data-communication barriers that block connected and automated vehicles in infrastructure-limited rural settings. It evaluates vehicular communication technologies and their limits, examines V2X architectures and protocols for coverage, reliability, and security, and uses text mining and topic modeling to surface current trends and research gaps for rural intelligent transportation systems.

Published in IEEE Access 2025; maps the rural V2X landscape and the connectivity, reliability, and security gaps that remain

keywords: V2X, DSRC, C-V2X, rural connectivity, vehicular networks

Agentic-UAV Control over Lossy Links

DroneComUAV / aerial autonomy

Moves the autonomy story from the ground to the air. It measures the network and safety cost of controlling a UAV with an LLM agent over a lossy communication link, where command drops and delay are the norm, and applies BlackWidow-style action masking so unsafe commands are filtered before they reach the aircraft.

Quantifies how link loss degrades agentic control, and keeps a hard safety gate between the planner and the vehicle

keywords: UAV, drone command and control, 5G, link loss, safe autonomy

Mentoring. This line is where my co-advising sits: Mahsa Tavasoli leads the rural-CAV intrusion-detection and communication work, Mohammad Sepahi co-authored the VIN privacy work, and several pieces are joint with the cybersecurity group at North Carolina A&T and Old Dominion University.