Abstract High-precision classification of encrypted traffic plays an important role in ensuring the reliability and safety of intelligent connected vehicles. However, the communication environment of vehicles is affected by complex traffic scenarios and changing external environments, which introduces noise into the observed traffic (e.g., padding artifacts and retransmission bursts). In addition, there is a strong similarity between different service categories. Therefore, existing encrypted traffic classification techniques are not applicable. To overcome these challenges, we propose SpACNet, a collaborative CNN-Transformer dual-path spectrum sensing classification network. Specifically, in addition to using stream sequence information, SpACNet also uses layered multi-scale spectrum recalibration technology and gated axial self-attention mechanism for frequency-domain information to suppress the influence of aliasing artifacts and noise. In terms of feature fusion, orthogonal constrained dynamic tensors and gating mechanisms are used to integrate and balance time-domain, frequency-domain tensors, and interaction tensors. We evaluate SpACNet and three advanced baseline methods based on public and real-world datasets. The results show that SpACNet outperforms existing methods and demonstrates robust performance on datasets containing highly similar traffic categories. In addition, a series of ablation experiments is conducted to demonstrate the advanced nature of the proposed method.
Wei et al. (2026) studied this question.