A Deep-Learning Based Cascading Framework for Intrusion Detection and Attack Classification on V2X-Based Autonomous Vehi

Overview

Vehicular Ad-Hoc Networks (VANETs) facilitate communication among autonomous vehicles (AVs) via V2X (Vehicle-to-Everything), making these AVs susceptible to various cyberattacks, including denial-of-service and data manipulation. To tackle this issue, previous research has developed intrusion detection models. However, to the best of our knowledge, these models may still train on messages sent by vehicles present in both training and testing, leading to unrealistic pattern recognition that reduces the framework's effectiveness when deployed. Thus, this work proposes a two-stage approach to address this issue: the first is a binary detection model classifying packet windows (sequences) as malicious or benign, while the second part classifies the attack type of malicious sequences. The Stage 1 model was trained on 2.2 million messages split by vehicle to prevent data leakage, a precaution consistent with Stage 2. In total, Stage 1 used 24 engineered features derived from position, velocity, and acceleration to develop a lightweight LSTM model providing 96.26% accuracy and an ROC-AUC of 97.6%. This model was further optimized for sequential processing to become viable for real-world scenarios, achieving an accuracy of 84.58% and an ROC-AUC of 96.2%. Stage 2, a Transformer classifying 19 attack types, improved a kinematic baseline of 70.81% to 88.23% through 16 engineered vehicle-independent features that use physics-based logic to generalize across unseen vehicles. The outcome of this work is a deployable framework for cities and industries to integrate into their infrastructure, enabling the growth of V2X-based AVs and ensuring passenger safety on the road.

Competition history

• CSEF 2026 Computational Science (Senior Division) · Entry S-07-36

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Computational Sciences projects
• CSEF Computational Sciences projects
• Every CSEF 2026 project