Intrusion Detection Based on Federated Context-Aware Embedded Deep Transfer Learning in Heterogeneous Networks

With the continuous advancement of the Internet of Everything paradigm, network intrusion detection systems (IDS) are confronted with multiple challenges in heterogeneous data sharing, integration, and security. To address these issues, this paper proposes a theoretical framework based on Word Embedded Federated Deep Transfer Learning (WE-FDTL) under the dual constraints of inconsistent data distribution and privacy preservation in heterogeneous network environments. The core innovation lies in establishing a latent vector-driven federated semantic aggregation mechanism to achieve cross-domain distributed representation alignment and knowledge fusion. At the representation learning level, a semantic space mapping model is constructed using contextual word embedding techniques to transform discrete heterogeneous network sequences into continuous dense vectors. This approach eliminates the need for explicit data standardization while preserving structural information in high-dimensional space and ensuring embedded privacy protection. For feature alignment, we propose a domain adaptation method based on latent space projection. By establishing latent space alignment constraints across participants, this method achieves cross-domain feature alignment and unified representation, while enabling effective adaptation of latent space matrices to mainstream deep learning models, thereby addressing the domain shift problem in heterogeneous federated transfer learning. At the federated optimization level, a semantically embedded federated aggregation mechanism is designed to facilitate global cross-domain knowledge sharing and integration through gradient transmission in latent vector space instead of raw data exchange. This framework ensures data privacy at terminal devices while maintaining effective knowledge fusion. In simulation experiments, six deep learning models were employed to evaluate multi-class classification performance across three scenarios: single-source domain deep learning, single-source domain federated deep learning, and cross-source domain WE-FDTL. Experimental results on the mixed NSL-KDD and UNSW-NB15 datasets demonstrate that WE-FDTL achieves a client-side training accuracy of 94.88% and a validation accuracy of 90.36%, confirming the theoretical effectiveness and practical advantages of the proposed IDS approach in heterogeneous network environments.