Malicious encrypted traffic detection method for service chains in multi-cloud environments

Abstract To resolve the pain points of insufficient feature characterization, weak cross-domain generalization and data privacy constraints in malicious encrypted traffic detection for multi-cloud service chains, this paper proposes a hierarchical multimodal detection framework (HMM-CSCD) tailored for encrypted traffic in multi-cloud service chain environments. The framework introduces three complementary feature modalities and employs federated self-supervised learning to learn generic traffic representations without exposing raw data across cloud domains. A dynamic attention-based fusion mechanism further enables real-time adaptation of modality weights according to instantaneous transmission states. Experiments conducted on the cloud platform of the China Electric Power Research Institute demonstrate that HMM-CSCD achieves consistent improvements in detection accuracy and cross-cloud generalization while maintaining an efficient single-flow inference latency of 2.32 ms. The proposed model offers a practical and privacy-preserving solution for securing network traffic in distributed cloud service chains.