Large Language Model (LLM) Based Resilient Intelligent Transport Systems

Abstract

Intelligent transport systems are increasingly dependent on interconnected devices and vehicular communications, making them vulnerable to reconnaissance attacks that precede large-scale intrusions. Traditional intrusion detection approaches often struggle with the scalability, redundancy of features, and complexity of dynamic traffic environments. To address these challenges, this paper introduces an LSH-Based Sparse Attention LLM Framework for reconnaissance attack detection. The framework applies Saint-Bowerbird Optimization (SBO) to select 23 optimal features from 41, ensuring efficiency and reducing redundancy. Categorical embeddings with drift analysis validate meaningful representation learning, while LSH-based sparse attention focuses on critical feature interactions with reduced complexity. The experimental results show that the model achieves 99.99% precision, outperforming RNN, LSTM, GRU and state-of-the-art models, confirming its robustness to secure intelligent transport systems.