← back

Research project

V2X Intrusion Detection for Autonomous Vehicles

Nov 2025 – Apr 2026

Independent research, with Mihir Bapat

  • Python
  • LSTM
  • Transformer
  • scikit-learn
  • AutoGluon

paper

highlights.md
  • Designed a two-stage framework: an LSTM that flags malicious V2X message sequences (96.26% accuracy, ROC-AUC 0.976) and a Transformer that classifies 19 attack types (88.23% accuracy, 88.05% F1, up from a 70.81% baseline) using physics-based, vehicle-independent features.
  • Split all data by vehicle so no vehicle appears in both training and testing, and built a sequential processor that classifies messages as they arrive (93.93% recall), with a ~142 KB model sized for vehicle onboard units.
  • Co-first author, IEEE ICCNCT 2026 (pp. 273–278). 1st Award, Greater San Diego Science and Engineering Fair; qualifier, California Science & Engineering Fair (2026).

Autonomous vehicles are starting to talk to each other and to the road around them over V2X (vehicle-to-everything) messages: I’m here, I’m going this fast, I’m braking. If one of those messages is forged, the car on the receiving end acts on a lie. After COSMOS got me interested in autonomous systems, my friend Mihir Bapat and I spent 11th grade building a system to catch those lies. Nobody mentored the project. We ran it ourselves, took it to the science fair, then wrote it up as a paper.

Two stages

We trained on the VeReMi Extension dataset, 3.2 million messages from 7,399 vehicles covering 19 attack types.

  • Stage 1: is this malicious? An LSTM looks at short windows of each vehicle’s messages and flags the suspicious ones, using 24 features built from position, velocity and acceleration. 96.26% accuracy, ROC-AUC 0.976.
  • Stage 2: what kind of attack? A Transformer classifies flagged traffic into the 19 attack types using physics-based features that don’t depend on which vehicle sent them, such as jerk and acceleration over time. 88.23% accuracy and 88.05% F1, up from a 70.81% baseline Transformer.

Keeping ourselves honest

The easiest way to get a great score on this dataset is to let the model memorize individual cars. We split every dataset by vehicle, so no vehicle appears in both training and testing, and the Stage 2 results come from 1,480 vehicles the model had never seen.

A car can’t wait for a full batch of messages, either, so we built a sequential processor that classifies each message as it arrives using a rolling 5-message buffer per vehicle. We tuned it for recall (93.93%), because in a moving car a missed attack is much worse than a false alarm. The Stage 1 model is about 142 KB, small enough for a vehicle’s onboard unit.

Where it went

  • 1st Award, Computer Science category, Greater San Diego Science and Engineering Fair (2026), and a qualifier for the California Science & Engineering Fair.
  • Published as A Deep-Learning based Cascading Framework for Intrusion Detection and Attack Classification on V2X-based Autonomous Vehicles at IEEE ICCNCT 2026, with Mihir and me as equal-contribution co-first authors (doi:10.1109/ICCNCT68477.2026.11590071).

↑↓ navigate · ↵ select · esc close