Scholarlink

Connected and autonomous vehicles (CAVs) promise improved safety, efficiency, and convenience but are vulnerable to adversarial attacks due to complex networks. This paper introduces a deep learning (DL) methodology integrating graph neural networks (GNNs) and Variational Autoencoders (VAEs). GNNs capture network relationships, learning robust node embeddings, while VAEs model latent space and reconstruction errors for probabilistic anomaly detection. This approach creates a graph from CAV data, extracts feature with GNNs, and uses VAEs to detect anomalies, forming a robust, updatable architecture to counter emerging threats and mitigate attack vectors. Evaluated on car-hacking datasets, the proposed method improves precision by 3.43%, recall by 2.94%, F-measure by 2.9%, training accuracy by 0.78%, and testing accuracy by 1.09%, with lower model loss, ensuring safer and more resilient autonomous vehicle operations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Adversarial Attack Detection in Connected and Autonomous Vehicle Networks Using GNN and Variational Autoencoders

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sireesha Prathigadapa,
  • V. Rajakumareswaran,
  • T. Sounderrajan,
  • K. Sivakumar

摘要

Connected and autonomous vehicles (CAVs) promise improved safety, efficiency, and convenience but are vulnerable to adversarial attacks due to complex networks. This paper introduces a deep learning (DL) methodology integrating graph neural networks (GNNs) and Variational Autoencoders (VAEs). GNNs capture network relationships, learning robust node embeddings, while VAEs model latent space and reconstruction errors for probabilistic anomaly detection. This approach creates a graph from CAV data, extracts feature with GNNs, and uses VAEs to detect anomalies, forming a robust, updatable architecture to counter emerging threats and mitigate attack vectors. Evaluated on car-hacking datasets, the proposed method improves precision by 3.43%, recall by 2.94%, F-measure by 2.9%, training accuracy by 0.78%, and testing accuracy by 1.09%, with lower model loss, ensuring safer and more resilient autonomous vehicle operations.