A Spatio-Temporal Transformer-Based Approach for Network Attack Detection

With the proliferation of dynamic network environments such as the Internet of Things (IoT) and Vehicle-to-Everything (V2X) networks, challenges related to data localization and privacy leakage have become increasingly prominent. Efficient and privacy-preserving real-time network attack detection has emerged as a critical issue in the field of cybersecurity. Existing methods struggle to balance handling dynamic topology changes, resisting privacy attacks, and maintaining high detection accuracy. To address this, this paper proposes an attack detection framework that integrates adaptive spatio-temporal graph generation, spatio-temporal Transformer, and privacy-enhanced federated learning. The approach designs a dynamic adjacency matrix optimization mechanism to capture real-time communication relationships, incorporates gradient compression and feature decoupling strategies to effectively mitigate model degradation caused by differential privacy noi se, and introduces a deep reinforcement learning module to dynamically adjust defense strategies based on real-time traffic feedback. This enables efficient collaborative training while ensuring data remains localized. Experimental results demonstrate that the proposed method achieves accuracy rates of 93.5% and 92.0% on the UNSW-NB15 and BoT-IoT datasets, respectively. Specifically, the proposed method achieves a 3.5 percentage point improvement over the strongest privacy-preserving baseline (DP-GNN) on the BoT-IoT dataset. Moreover, under the same privacy budget, it significantly reduces the success rate of membership inference attacks. This work provides novel insights into distributed secure learning in high-privacy-risk dynamic environments, with its low-latency and high-robustness characteristics supporting real-time threat perception and privacy-secure responses on edge devices.