TAGT: A Time-Aware Graph Transformer for Securing Time-Sensitive Networking

Le résumé fourni par la source

The precision of time synchronization is the cornerstone of safety-critical applications in Automotive Ethernet and Time-Sensitive Networking (TSN). The IEEE 802.1AS protocol (gPTP), while widely adopted, lacks built-in security mechanisms, making it vulnerable to sophisticated delay attacks and packet manipulation. Existing Deep Learning-based Intrusion Detection Systems (IDS) often treat network traffic as discrete sequences, ignoring the continuous physical time dynamics and the complex spatial topological dependencies inherent in synchronization protocols.To address these limitations, this study proposes TAGT, a novel Time-Aware Graph Transformer framework for robust gPTP attack detection. Unlike standard Transformers that rely on discrete positional encodings, TAGT incorporates a Learnable Continuous Time-Encoding (LCTE) mechanism. LCTE maps irregular packet arrival intervals into high-dimensional frequency representations, enabling the model to capture microsecond-level clock jitter and drift anomalies. Furthermore, we integrate Graph Attention Networks (GAT) to extract spatial features from the network topology, fusing them with temporal embeddings to learn a unified spatio-temporal representation. Extensive experiments on high-fidelity OMNeT++ simulations demonstrate that TAGT achieves an F1-score of 95.65%, outperforming existing baseline models (e.g., LSTM, GCN) by a significant margin. The results validate that our approach is highly effective in detecting low-frequency attacks and is robust to topological changes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.